xref: /linux/kernel/bpf/arraymap.c (revision f6f3bac08ff9855d803081a353a1fafaa8845739)
1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2  * Copyright (c) 2016,2017 Facebook
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of version 2 of the GNU General Public
6  * License as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful, but
9  * WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11  * General Public License for more details.
12  */
13 #include <linux/bpf.h>
14 #include <linux/btf.h>
15 #include <linux/err.h>
16 #include <linux/slab.h>
17 #include <linux/mm.h>
18 #include <linux/filter.h>
19 #include <linux/perf_event.h>
20 #include <uapi/linux/btf.h>
21 
22 #include "map_in_map.h"
23 
24 #define ARRAY_CREATE_FLAG_MASK \
25 	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
26 
27 static void bpf_array_free_percpu(struct bpf_array *array)
28 {
29 	int i;
30 
31 	for (i = 0; i < array->map.max_entries; i++) {
32 		free_percpu(array->pptrs[i]);
33 		cond_resched();
34 	}
35 }
36 
37 static int bpf_array_alloc_percpu(struct bpf_array *array)
38 {
39 	void __percpu *ptr;
40 	int i;
41 
42 	for (i = 0; i < array->map.max_entries; i++) {
43 		ptr = __alloc_percpu_gfp(array->elem_size, 8,
44 					 GFP_USER | __GFP_NOWARN);
45 		if (!ptr) {
46 			bpf_array_free_percpu(array);
47 			return -ENOMEM;
48 		}
49 		array->pptrs[i] = ptr;
50 		cond_resched();
51 	}
52 
53 	return 0;
54 }
55 
56 /* Called from syscall */
57 int array_map_alloc_check(union bpf_attr *attr)
58 {
59 	bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
60 	int numa_node = bpf_map_attr_numa_node(attr);
61 
62 	/* check sanity of attributes */
63 	if (attr->max_entries == 0 || attr->key_size != 4 ||
64 	    attr->value_size == 0 ||
65 	    attr->map_flags & ~ARRAY_CREATE_FLAG_MASK ||
66 	    (percpu && numa_node != NUMA_NO_NODE))
67 		return -EINVAL;
68 
69 	if (attr->value_size > KMALLOC_MAX_SIZE)
70 		/* if value_size is bigger, the user space won't be able to
71 		 * access the elements.
72 		 */
73 		return -E2BIG;
74 
75 	return 0;
76 }
77 
78 static struct bpf_map *array_map_alloc(union bpf_attr *attr)
79 {
80 	bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
81 	int ret, numa_node = bpf_map_attr_numa_node(attr);
82 	u32 elem_size, index_mask, max_entries;
83 	bool unpriv = !capable(CAP_SYS_ADMIN);
84 	u64 cost, array_size, mask64;
85 	struct bpf_array *array;
86 
87 	elem_size = round_up(attr->value_size, 8);
88 
89 	max_entries = attr->max_entries;
90 
91 	/* On 32 bit archs roundup_pow_of_two() with max_entries that has
92 	 * upper most bit set in u32 space is undefined behavior due to
93 	 * resulting 1U << 32, so do it manually here in u64 space.
94 	 */
95 	mask64 = fls_long(max_entries - 1);
96 	mask64 = 1ULL << mask64;
97 	mask64 -= 1;
98 
99 	index_mask = mask64;
100 	if (unpriv) {
101 		/* round up array size to nearest power of 2,
102 		 * since cpu will speculate within index_mask limits
103 		 */
104 		max_entries = index_mask + 1;
105 		/* Check for overflows. */
106 		if (max_entries < attr->max_entries)
107 			return ERR_PTR(-E2BIG);
108 	}
109 
110 	array_size = sizeof(*array);
111 	if (percpu)
112 		array_size += (u64) max_entries * sizeof(void *);
113 	else
114 		array_size += (u64) max_entries * elem_size;
115 
116 	/* make sure there is no u32 overflow later in round_up() */
117 	cost = array_size;
118 	if (cost >= U32_MAX - PAGE_SIZE)
119 		return ERR_PTR(-ENOMEM);
120 	if (percpu) {
121 		cost += (u64)attr->max_entries * elem_size * num_possible_cpus();
122 		if (cost >= U32_MAX - PAGE_SIZE)
123 			return ERR_PTR(-ENOMEM);
124 	}
125 	cost = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
126 
127 	ret = bpf_map_precharge_memlock(cost);
128 	if (ret < 0)
129 		return ERR_PTR(ret);
130 
131 	/* allocate all map elements and zero-initialize them */
132 	array = bpf_map_area_alloc(array_size, numa_node);
133 	if (!array)
134 		return ERR_PTR(-ENOMEM);
135 	array->index_mask = index_mask;
136 	array->map.unpriv_array = unpriv;
137 
138 	/* copy mandatory map attributes */
139 	bpf_map_init_from_attr(&array->map, attr);
140 	array->map.pages = cost;
141 	array->elem_size = elem_size;
142 
143 	if (percpu && bpf_array_alloc_percpu(array)) {
144 		bpf_map_area_free(array);
145 		return ERR_PTR(-ENOMEM);
146 	}
147 
148 	return &array->map;
149 }
150 
151 /* Called from syscall or from eBPF program */
152 static void *array_map_lookup_elem(struct bpf_map *map, void *key)
153 {
154 	struct bpf_array *array = container_of(map, struct bpf_array, map);
155 	u32 index = *(u32 *)key;
156 
157 	if (unlikely(index >= array->map.max_entries))
158 		return NULL;
159 
160 	return array->value + array->elem_size * (index & array->index_mask);
161 }
162 
163 /* emit BPF instructions equivalent to C code of array_map_lookup_elem() */
164 static u32 array_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
165 {
166 	struct bpf_array *array = container_of(map, struct bpf_array, map);
167 	struct bpf_insn *insn = insn_buf;
168 	u32 elem_size = round_up(map->value_size, 8);
169 	const int ret = BPF_REG_0;
170 	const int map_ptr = BPF_REG_1;
171 	const int index = BPF_REG_2;
172 
173 	*insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
174 	*insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
175 	if (map->unpriv_array) {
176 		*insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 4);
177 		*insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
178 	} else {
179 		*insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 3);
180 	}
181 
182 	if (is_power_of_2(elem_size)) {
183 		*insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
184 	} else {
185 		*insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
186 	}
187 	*insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
188 	*insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
189 	*insn++ = BPF_MOV64_IMM(ret, 0);
190 	return insn - insn_buf;
191 }
192 
193 /* Called from eBPF program */
194 static void *percpu_array_map_lookup_elem(struct bpf_map *map, void *key)
195 {
196 	struct bpf_array *array = container_of(map, struct bpf_array, map);
197 	u32 index = *(u32 *)key;
198 
199 	if (unlikely(index >= array->map.max_entries))
200 		return NULL;
201 
202 	return this_cpu_ptr(array->pptrs[index & array->index_mask]);
203 }
204 
205 int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value)
206 {
207 	struct bpf_array *array = container_of(map, struct bpf_array, map);
208 	u32 index = *(u32 *)key;
209 	void __percpu *pptr;
210 	int cpu, off = 0;
211 	u32 size;
212 
213 	if (unlikely(index >= array->map.max_entries))
214 		return -ENOENT;
215 
216 	/* per_cpu areas are zero-filled and bpf programs can only
217 	 * access 'value_size' of them, so copying rounded areas
218 	 * will not leak any kernel data
219 	 */
220 	size = round_up(map->value_size, 8);
221 	rcu_read_lock();
222 	pptr = array->pptrs[index & array->index_mask];
223 	for_each_possible_cpu(cpu) {
224 		bpf_long_memcpy(value + off, per_cpu_ptr(pptr, cpu), size);
225 		off += size;
226 	}
227 	rcu_read_unlock();
228 	return 0;
229 }
230 
231 /* Called from syscall */
232 static int array_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
233 {
234 	struct bpf_array *array = container_of(map, struct bpf_array, map);
235 	u32 index = key ? *(u32 *)key : U32_MAX;
236 	u32 *next = (u32 *)next_key;
237 
238 	if (index >= array->map.max_entries) {
239 		*next = 0;
240 		return 0;
241 	}
242 
243 	if (index == array->map.max_entries - 1)
244 		return -ENOENT;
245 
246 	*next = index + 1;
247 	return 0;
248 }
249 
250 /* Called from syscall or from eBPF program */
251 static int array_map_update_elem(struct bpf_map *map, void *key, void *value,
252 				 u64 map_flags)
253 {
254 	struct bpf_array *array = container_of(map, struct bpf_array, map);
255 	u32 index = *(u32 *)key;
256 
257 	if (unlikely(map_flags > BPF_EXIST))
258 		/* unknown flags */
259 		return -EINVAL;
260 
261 	if (unlikely(index >= array->map.max_entries))
262 		/* all elements were pre-allocated, cannot insert a new one */
263 		return -E2BIG;
264 
265 	if (unlikely(map_flags == BPF_NOEXIST))
266 		/* all elements already exist */
267 		return -EEXIST;
268 
269 	if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
270 		memcpy(this_cpu_ptr(array->pptrs[index & array->index_mask]),
271 		       value, map->value_size);
272 	else
273 		memcpy(array->value +
274 		       array->elem_size * (index & array->index_mask),
275 		       value, map->value_size);
276 	return 0;
277 }
278 
279 int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
280 			    u64 map_flags)
281 {
282 	struct bpf_array *array = container_of(map, struct bpf_array, map);
283 	u32 index = *(u32 *)key;
284 	void __percpu *pptr;
285 	int cpu, off = 0;
286 	u32 size;
287 
288 	if (unlikely(map_flags > BPF_EXIST))
289 		/* unknown flags */
290 		return -EINVAL;
291 
292 	if (unlikely(index >= array->map.max_entries))
293 		/* all elements were pre-allocated, cannot insert a new one */
294 		return -E2BIG;
295 
296 	if (unlikely(map_flags == BPF_NOEXIST))
297 		/* all elements already exist */
298 		return -EEXIST;
299 
300 	/* the user space will provide round_up(value_size, 8) bytes that
301 	 * will be copied into per-cpu area. bpf programs can only access
302 	 * value_size of it. During lookup the same extra bytes will be
303 	 * returned or zeros which were zero-filled by percpu_alloc,
304 	 * so no kernel data leaks possible
305 	 */
306 	size = round_up(map->value_size, 8);
307 	rcu_read_lock();
308 	pptr = array->pptrs[index & array->index_mask];
309 	for_each_possible_cpu(cpu) {
310 		bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value + off, size);
311 		off += size;
312 	}
313 	rcu_read_unlock();
314 	return 0;
315 }
316 
317 /* Called from syscall or from eBPF program */
318 static int array_map_delete_elem(struct bpf_map *map, void *key)
319 {
320 	return -EINVAL;
321 }
322 
323 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
324 static void array_map_free(struct bpf_map *map)
325 {
326 	struct bpf_array *array = container_of(map, struct bpf_array, map);
327 
328 	/* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
329 	 * so the programs (can be more than one that used this map) were
330 	 * disconnected from events. Wait for outstanding programs to complete
331 	 * and free the array
332 	 */
333 	synchronize_rcu();
334 
335 	if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
336 		bpf_array_free_percpu(array);
337 
338 	bpf_map_area_free(array);
339 }
340 
341 static void array_map_seq_show_elem(struct bpf_map *map, void *key,
342 				    struct seq_file *m)
343 {
344 	void *value;
345 
346 	rcu_read_lock();
347 
348 	value = array_map_lookup_elem(map, key);
349 	if (!value) {
350 		rcu_read_unlock();
351 		return;
352 	}
353 
354 	seq_printf(m, "%u: ", *(u32 *)key);
355 	btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
356 	seq_puts(m, "\n");
357 
358 	rcu_read_unlock();
359 }
360 
361 static void percpu_array_map_seq_show_elem(struct bpf_map *map, void *key,
362 					   struct seq_file *m)
363 {
364 	struct bpf_array *array = container_of(map, struct bpf_array, map);
365 	u32 index = *(u32 *)key;
366 	void __percpu *pptr;
367 	int cpu;
368 
369 	rcu_read_lock();
370 
371 	seq_printf(m, "%u: {\n", *(u32 *)key);
372 	pptr = array->pptrs[index & array->index_mask];
373 	for_each_possible_cpu(cpu) {
374 		seq_printf(m, "\tcpu%d: ", cpu);
375 		btf_type_seq_show(map->btf, map->btf_value_type_id,
376 				  per_cpu_ptr(pptr, cpu), m);
377 		seq_puts(m, "\n");
378 	}
379 	seq_puts(m, "}\n");
380 
381 	rcu_read_unlock();
382 }
383 
384 static int array_map_check_btf(const struct bpf_map *map,
385 			       const struct btf_type *key_type,
386 			       const struct btf_type *value_type)
387 {
388 	u32 int_data;
389 
390 	if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
391 		return -EINVAL;
392 
393 	int_data = *(u32 *)(key_type + 1);
394 	/* bpf array can only take a u32 key. This check makes sure
395 	 * that the btf matches the attr used during map_create.
396 	 */
397 	if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
398 		return -EINVAL;
399 
400 	return 0;
401 }
402 
403 const struct bpf_map_ops array_map_ops = {
404 	.map_alloc_check = array_map_alloc_check,
405 	.map_alloc = array_map_alloc,
406 	.map_free = array_map_free,
407 	.map_get_next_key = array_map_get_next_key,
408 	.map_lookup_elem = array_map_lookup_elem,
409 	.map_update_elem = array_map_update_elem,
410 	.map_delete_elem = array_map_delete_elem,
411 	.map_gen_lookup = array_map_gen_lookup,
412 	.map_seq_show_elem = array_map_seq_show_elem,
413 	.map_check_btf = array_map_check_btf,
414 };
415 
416 const struct bpf_map_ops percpu_array_map_ops = {
417 	.map_alloc_check = array_map_alloc_check,
418 	.map_alloc = array_map_alloc,
419 	.map_free = array_map_free,
420 	.map_get_next_key = array_map_get_next_key,
421 	.map_lookup_elem = percpu_array_map_lookup_elem,
422 	.map_update_elem = array_map_update_elem,
423 	.map_delete_elem = array_map_delete_elem,
424 	.map_seq_show_elem = percpu_array_map_seq_show_elem,
425 	.map_check_btf = array_map_check_btf,
426 };
427 
428 static int fd_array_map_alloc_check(union bpf_attr *attr)
429 {
430 	/* only file descriptors can be stored in this type of map */
431 	if (attr->value_size != sizeof(u32))
432 		return -EINVAL;
433 	return array_map_alloc_check(attr);
434 }
435 
436 static void fd_array_map_free(struct bpf_map *map)
437 {
438 	struct bpf_array *array = container_of(map, struct bpf_array, map);
439 	int i;
440 
441 	synchronize_rcu();
442 
443 	/* make sure it's empty */
444 	for (i = 0; i < array->map.max_entries; i++)
445 		BUG_ON(array->ptrs[i] != NULL);
446 
447 	bpf_map_area_free(array);
448 }
449 
450 static void *fd_array_map_lookup_elem(struct bpf_map *map, void *key)
451 {
452 	return NULL;
453 }
454 
455 /* only called from syscall */
456 int bpf_fd_array_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
457 {
458 	void **elem, *ptr;
459 	int ret =  0;
460 
461 	if (!map->ops->map_fd_sys_lookup_elem)
462 		return -ENOTSUPP;
463 
464 	rcu_read_lock();
465 	elem = array_map_lookup_elem(map, key);
466 	if (elem && (ptr = READ_ONCE(*elem)))
467 		*value = map->ops->map_fd_sys_lookup_elem(ptr);
468 	else
469 		ret = -ENOENT;
470 	rcu_read_unlock();
471 
472 	return ret;
473 }
474 
475 /* only called from syscall */
476 int bpf_fd_array_map_update_elem(struct bpf_map *map, struct file *map_file,
477 				 void *key, void *value, u64 map_flags)
478 {
479 	struct bpf_array *array = container_of(map, struct bpf_array, map);
480 	void *new_ptr, *old_ptr;
481 	u32 index = *(u32 *)key, ufd;
482 
483 	if (map_flags != BPF_ANY)
484 		return -EINVAL;
485 
486 	if (index >= array->map.max_entries)
487 		return -E2BIG;
488 
489 	ufd = *(u32 *)value;
490 	new_ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
491 	if (IS_ERR(new_ptr))
492 		return PTR_ERR(new_ptr);
493 
494 	old_ptr = xchg(array->ptrs + index, new_ptr);
495 	if (old_ptr)
496 		map->ops->map_fd_put_ptr(old_ptr);
497 
498 	return 0;
499 }
500 
501 static int fd_array_map_delete_elem(struct bpf_map *map, void *key)
502 {
503 	struct bpf_array *array = container_of(map, struct bpf_array, map);
504 	void *old_ptr;
505 	u32 index = *(u32 *)key;
506 
507 	if (index >= array->map.max_entries)
508 		return -E2BIG;
509 
510 	old_ptr = xchg(array->ptrs + index, NULL);
511 	if (old_ptr) {
512 		map->ops->map_fd_put_ptr(old_ptr);
513 		return 0;
514 	} else {
515 		return -ENOENT;
516 	}
517 }
518 
519 static void *prog_fd_array_get_ptr(struct bpf_map *map,
520 				   struct file *map_file, int fd)
521 {
522 	struct bpf_array *array = container_of(map, struct bpf_array, map);
523 	struct bpf_prog *prog = bpf_prog_get(fd);
524 
525 	if (IS_ERR(prog))
526 		return prog;
527 
528 	if (!bpf_prog_array_compatible(array, prog)) {
529 		bpf_prog_put(prog);
530 		return ERR_PTR(-EINVAL);
531 	}
532 
533 	return prog;
534 }
535 
536 static void prog_fd_array_put_ptr(void *ptr)
537 {
538 	bpf_prog_put(ptr);
539 }
540 
541 static u32 prog_fd_array_sys_lookup_elem(void *ptr)
542 {
543 	return ((struct bpf_prog *)ptr)->aux->id;
544 }
545 
546 /* decrement refcnt of all bpf_progs that are stored in this map */
547 static void bpf_fd_array_map_clear(struct bpf_map *map)
548 {
549 	struct bpf_array *array = container_of(map, struct bpf_array, map);
550 	int i;
551 
552 	for (i = 0; i < array->map.max_entries; i++)
553 		fd_array_map_delete_elem(map, &i);
554 }
555 
556 const struct bpf_map_ops prog_array_map_ops = {
557 	.map_alloc_check = fd_array_map_alloc_check,
558 	.map_alloc = array_map_alloc,
559 	.map_free = fd_array_map_free,
560 	.map_get_next_key = array_map_get_next_key,
561 	.map_lookup_elem = fd_array_map_lookup_elem,
562 	.map_delete_elem = fd_array_map_delete_elem,
563 	.map_fd_get_ptr = prog_fd_array_get_ptr,
564 	.map_fd_put_ptr = prog_fd_array_put_ptr,
565 	.map_fd_sys_lookup_elem = prog_fd_array_sys_lookup_elem,
566 	.map_release_uref = bpf_fd_array_map_clear,
567 	.map_check_btf = map_check_no_btf,
568 };
569 
570 static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
571 						   struct file *map_file)
572 {
573 	struct bpf_event_entry *ee;
574 
575 	ee = kzalloc(sizeof(*ee), GFP_ATOMIC);
576 	if (ee) {
577 		ee->event = perf_file->private_data;
578 		ee->perf_file = perf_file;
579 		ee->map_file = map_file;
580 	}
581 
582 	return ee;
583 }
584 
585 static void __bpf_event_entry_free(struct rcu_head *rcu)
586 {
587 	struct bpf_event_entry *ee;
588 
589 	ee = container_of(rcu, struct bpf_event_entry, rcu);
590 	fput(ee->perf_file);
591 	kfree(ee);
592 }
593 
594 static void bpf_event_entry_free_rcu(struct bpf_event_entry *ee)
595 {
596 	call_rcu(&ee->rcu, __bpf_event_entry_free);
597 }
598 
599 static void *perf_event_fd_array_get_ptr(struct bpf_map *map,
600 					 struct file *map_file, int fd)
601 {
602 	struct bpf_event_entry *ee;
603 	struct perf_event *event;
604 	struct file *perf_file;
605 	u64 value;
606 
607 	perf_file = perf_event_get(fd);
608 	if (IS_ERR(perf_file))
609 		return perf_file;
610 
611 	ee = ERR_PTR(-EOPNOTSUPP);
612 	event = perf_file->private_data;
613 	if (perf_event_read_local(event, &value, NULL, NULL) == -EOPNOTSUPP)
614 		goto err_out;
615 
616 	ee = bpf_event_entry_gen(perf_file, map_file);
617 	if (ee)
618 		return ee;
619 	ee = ERR_PTR(-ENOMEM);
620 err_out:
621 	fput(perf_file);
622 	return ee;
623 }
624 
625 static void perf_event_fd_array_put_ptr(void *ptr)
626 {
627 	bpf_event_entry_free_rcu(ptr);
628 }
629 
630 static void perf_event_fd_array_release(struct bpf_map *map,
631 					struct file *map_file)
632 {
633 	struct bpf_array *array = container_of(map, struct bpf_array, map);
634 	struct bpf_event_entry *ee;
635 	int i;
636 
637 	rcu_read_lock();
638 	for (i = 0; i < array->map.max_entries; i++) {
639 		ee = READ_ONCE(array->ptrs[i]);
640 		if (ee && ee->map_file == map_file)
641 			fd_array_map_delete_elem(map, &i);
642 	}
643 	rcu_read_unlock();
644 }
645 
646 const struct bpf_map_ops perf_event_array_map_ops = {
647 	.map_alloc_check = fd_array_map_alloc_check,
648 	.map_alloc = array_map_alloc,
649 	.map_free = fd_array_map_free,
650 	.map_get_next_key = array_map_get_next_key,
651 	.map_lookup_elem = fd_array_map_lookup_elem,
652 	.map_delete_elem = fd_array_map_delete_elem,
653 	.map_fd_get_ptr = perf_event_fd_array_get_ptr,
654 	.map_fd_put_ptr = perf_event_fd_array_put_ptr,
655 	.map_release = perf_event_fd_array_release,
656 	.map_check_btf = map_check_no_btf,
657 };
658 
659 #ifdef CONFIG_CGROUPS
660 static void *cgroup_fd_array_get_ptr(struct bpf_map *map,
661 				     struct file *map_file /* not used */,
662 				     int fd)
663 {
664 	return cgroup_get_from_fd(fd);
665 }
666 
667 static void cgroup_fd_array_put_ptr(void *ptr)
668 {
669 	/* cgroup_put free cgrp after a rcu grace period */
670 	cgroup_put(ptr);
671 }
672 
673 static void cgroup_fd_array_free(struct bpf_map *map)
674 {
675 	bpf_fd_array_map_clear(map);
676 	fd_array_map_free(map);
677 }
678 
679 const struct bpf_map_ops cgroup_array_map_ops = {
680 	.map_alloc_check = fd_array_map_alloc_check,
681 	.map_alloc = array_map_alloc,
682 	.map_free = cgroup_fd_array_free,
683 	.map_get_next_key = array_map_get_next_key,
684 	.map_lookup_elem = fd_array_map_lookup_elem,
685 	.map_delete_elem = fd_array_map_delete_elem,
686 	.map_fd_get_ptr = cgroup_fd_array_get_ptr,
687 	.map_fd_put_ptr = cgroup_fd_array_put_ptr,
688 	.map_check_btf = map_check_no_btf,
689 };
690 #endif
691 
692 static struct bpf_map *array_of_map_alloc(union bpf_attr *attr)
693 {
694 	struct bpf_map *map, *inner_map_meta;
695 
696 	inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
697 	if (IS_ERR(inner_map_meta))
698 		return inner_map_meta;
699 
700 	map = array_map_alloc(attr);
701 	if (IS_ERR(map)) {
702 		bpf_map_meta_free(inner_map_meta);
703 		return map;
704 	}
705 
706 	map->inner_map_meta = inner_map_meta;
707 
708 	return map;
709 }
710 
711 static void array_of_map_free(struct bpf_map *map)
712 {
713 	/* map->inner_map_meta is only accessed by syscall which
714 	 * is protected by fdget/fdput.
715 	 */
716 	bpf_map_meta_free(map->inner_map_meta);
717 	bpf_fd_array_map_clear(map);
718 	fd_array_map_free(map);
719 }
720 
721 static void *array_of_map_lookup_elem(struct bpf_map *map, void *key)
722 {
723 	struct bpf_map **inner_map = array_map_lookup_elem(map, key);
724 
725 	if (!inner_map)
726 		return NULL;
727 
728 	return READ_ONCE(*inner_map);
729 }
730 
731 static u32 array_of_map_gen_lookup(struct bpf_map *map,
732 				   struct bpf_insn *insn_buf)
733 {
734 	struct bpf_array *array = container_of(map, struct bpf_array, map);
735 	u32 elem_size = round_up(map->value_size, 8);
736 	struct bpf_insn *insn = insn_buf;
737 	const int ret = BPF_REG_0;
738 	const int map_ptr = BPF_REG_1;
739 	const int index = BPF_REG_2;
740 
741 	*insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
742 	*insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
743 	if (map->unpriv_array) {
744 		*insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 6);
745 		*insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
746 	} else {
747 		*insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
748 	}
749 	if (is_power_of_2(elem_size))
750 		*insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
751 	else
752 		*insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
753 	*insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
754 	*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
755 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
756 	*insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
757 	*insn++ = BPF_MOV64_IMM(ret, 0);
758 
759 	return insn - insn_buf;
760 }
761 
762 const struct bpf_map_ops array_of_maps_map_ops = {
763 	.map_alloc_check = fd_array_map_alloc_check,
764 	.map_alloc = array_of_map_alloc,
765 	.map_free = array_of_map_free,
766 	.map_get_next_key = array_map_get_next_key,
767 	.map_lookup_elem = array_of_map_lookup_elem,
768 	.map_delete_elem = fd_array_map_delete_elem,
769 	.map_fd_get_ptr = bpf_map_fd_get_ptr,
770 	.map_fd_put_ptr = bpf_map_fd_put_ptr,
771 	.map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
772 	.map_gen_lookup = array_of_map_gen_lookup,
773 	.map_check_btf = map_check_no_btf,
774 };
775