Loading...
1/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 *
3 * This program is free software; you can redistribute it and/or
4 * modify it under the terms of version 2 of the GNU General Public
5 * License as published by the Free Software Foundation.
6 *
7 * This program is distributed in the hope that it will be useful, but
8 * WITHOUT ANY WARRANTY; without even the implied warranty of
9 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
10 * General Public License for more details.
11 */
12#include <linux/bpf.h>
13#include <linux/err.h>
14#include <linux/vmalloc.h>
15#include <linux/slab.h>
16#include <linux/mm.h>
17#include <linux/filter.h>
18#include <linux/perf_event.h>
19
20static void bpf_array_free_percpu(struct bpf_array *array)
21{
22 int i;
23
24 for (i = 0; i < array->map.max_entries; i++)
25 free_percpu(array->pptrs[i]);
26}
27
28static int bpf_array_alloc_percpu(struct bpf_array *array)
29{
30 void __percpu *ptr;
31 int i;
32
33 for (i = 0; i < array->map.max_entries; i++) {
34 ptr = __alloc_percpu_gfp(array->elem_size, 8,
35 GFP_USER | __GFP_NOWARN);
36 if (!ptr) {
37 bpf_array_free_percpu(array);
38 return -ENOMEM;
39 }
40 array->pptrs[i] = ptr;
41 }
42
43 return 0;
44}
45
46/* Called from syscall */
47static struct bpf_map *array_map_alloc(union bpf_attr *attr)
48{
49 bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
50 struct bpf_array *array;
51 u64 array_size;
52 u32 elem_size;
53
54 /* check sanity of attributes */
55 if (attr->max_entries == 0 || attr->key_size != 4 ||
56 attr->value_size == 0 || attr->map_flags)
57 return ERR_PTR(-EINVAL);
58
59 if (attr->value_size >= 1 << (KMALLOC_SHIFT_MAX - 1))
60 /* if value_size is bigger, the user space won't be able to
61 * access the elements.
62 */
63 return ERR_PTR(-E2BIG);
64
65 elem_size = round_up(attr->value_size, 8);
66
67 array_size = sizeof(*array);
68 if (percpu)
69 array_size += (u64) attr->max_entries * sizeof(void *);
70 else
71 array_size += (u64) attr->max_entries * elem_size;
72
73 /* make sure there is no u32 overflow later in round_up() */
74 if (array_size >= U32_MAX - PAGE_SIZE)
75 return ERR_PTR(-ENOMEM);
76
77
78 /* allocate all map elements and zero-initialize them */
79 array = kzalloc(array_size, GFP_USER | __GFP_NOWARN);
80 if (!array) {
81 array = vzalloc(array_size);
82 if (!array)
83 return ERR_PTR(-ENOMEM);
84 }
85
86 /* copy mandatory map attributes */
87 array->map.map_type = attr->map_type;
88 array->map.key_size = attr->key_size;
89 array->map.value_size = attr->value_size;
90 array->map.max_entries = attr->max_entries;
91 array->elem_size = elem_size;
92
93 if (!percpu)
94 goto out;
95
96 array_size += (u64) attr->max_entries * elem_size * num_possible_cpus();
97
98 if (array_size >= U32_MAX - PAGE_SIZE ||
99 elem_size > PCPU_MIN_UNIT_SIZE || bpf_array_alloc_percpu(array)) {
100 kvfree(array);
101 return ERR_PTR(-ENOMEM);
102 }
103out:
104 array->map.pages = round_up(array_size, PAGE_SIZE) >> PAGE_SHIFT;
105
106 return &array->map;
107}
108
109/* Called from syscall or from eBPF program */
110static void *array_map_lookup_elem(struct bpf_map *map, void *key)
111{
112 struct bpf_array *array = container_of(map, struct bpf_array, map);
113 u32 index = *(u32 *)key;
114
115 if (unlikely(index >= array->map.max_entries))
116 return NULL;
117
118 return array->value + array->elem_size * index;
119}
120
121/* Called from eBPF program */
122static void *percpu_array_map_lookup_elem(struct bpf_map *map, void *key)
123{
124 struct bpf_array *array = container_of(map, struct bpf_array, map);
125 u32 index = *(u32 *)key;
126
127 if (unlikely(index >= array->map.max_entries))
128 return NULL;
129
130 return this_cpu_ptr(array->pptrs[index]);
131}
132
133int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value)
134{
135 struct bpf_array *array = container_of(map, struct bpf_array, map);
136 u32 index = *(u32 *)key;
137 void __percpu *pptr;
138 int cpu, off = 0;
139 u32 size;
140
141 if (unlikely(index >= array->map.max_entries))
142 return -ENOENT;
143
144 /* per_cpu areas are zero-filled and bpf programs can only
145 * access 'value_size' of them, so copying rounded areas
146 * will not leak any kernel data
147 */
148 size = round_up(map->value_size, 8);
149 rcu_read_lock();
150 pptr = array->pptrs[index];
151 for_each_possible_cpu(cpu) {
152 bpf_long_memcpy(value + off, per_cpu_ptr(pptr, cpu), size);
153 off += size;
154 }
155 rcu_read_unlock();
156 return 0;
157}
158
159/* Called from syscall */
160static int array_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
161{
162 struct bpf_array *array = container_of(map, struct bpf_array, map);
163 u32 index = *(u32 *)key;
164 u32 *next = (u32 *)next_key;
165
166 if (index >= array->map.max_entries) {
167 *next = 0;
168 return 0;
169 }
170
171 if (index == array->map.max_entries - 1)
172 return -ENOENT;
173
174 *next = index + 1;
175 return 0;
176}
177
178/* Called from syscall or from eBPF program */
179static int array_map_update_elem(struct bpf_map *map, void *key, void *value,
180 u64 map_flags)
181{
182 struct bpf_array *array = container_of(map, struct bpf_array, map);
183 u32 index = *(u32 *)key;
184
185 if (unlikely(map_flags > BPF_EXIST))
186 /* unknown flags */
187 return -EINVAL;
188
189 if (unlikely(index >= array->map.max_entries))
190 /* all elements were pre-allocated, cannot insert a new one */
191 return -E2BIG;
192
193 if (unlikely(map_flags == BPF_NOEXIST))
194 /* all elements already exist */
195 return -EEXIST;
196
197 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
198 memcpy(this_cpu_ptr(array->pptrs[index]),
199 value, map->value_size);
200 else
201 memcpy(array->value + array->elem_size * index,
202 value, map->value_size);
203 return 0;
204}
205
206int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
207 u64 map_flags)
208{
209 struct bpf_array *array = container_of(map, struct bpf_array, map);
210 u32 index = *(u32 *)key;
211 void __percpu *pptr;
212 int cpu, off = 0;
213 u32 size;
214
215 if (unlikely(map_flags > BPF_EXIST))
216 /* unknown flags */
217 return -EINVAL;
218
219 if (unlikely(index >= array->map.max_entries))
220 /* all elements were pre-allocated, cannot insert a new one */
221 return -E2BIG;
222
223 if (unlikely(map_flags == BPF_NOEXIST))
224 /* all elements already exist */
225 return -EEXIST;
226
227 /* the user space will provide round_up(value_size, 8) bytes that
228 * will be copied into per-cpu area. bpf programs can only access
229 * value_size of it. During lookup the same extra bytes will be
230 * returned or zeros which were zero-filled by percpu_alloc,
231 * so no kernel data leaks possible
232 */
233 size = round_up(map->value_size, 8);
234 rcu_read_lock();
235 pptr = array->pptrs[index];
236 for_each_possible_cpu(cpu) {
237 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value + off, size);
238 off += size;
239 }
240 rcu_read_unlock();
241 return 0;
242}
243
244/* Called from syscall or from eBPF program */
245static int array_map_delete_elem(struct bpf_map *map, void *key)
246{
247 return -EINVAL;
248}
249
250/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
251static void array_map_free(struct bpf_map *map)
252{
253 struct bpf_array *array = container_of(map, struct bpf_array, map);
254
255 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
256 * so the programs (can be more than one that used this map) were
257 * disconnected from events. Wait for outstanding programs to complete
258 * and free the array
259 */
260 synchronize_rcu();
261
262 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
263 bpf_array_free_percpu(array);
264
265 kvfree(array);
266}
267
268static const struct bpf_map_ops array_ops = {
269 .map_alloc = array_map_alloc,
270 .map_free = array_map_free,
271 .map_get_next_key = array_map_get_next_key,
272 .map_lookup_elem = array_map_lookup_elem,
273 .map_update_elem = array_map_update_elem,
274 .map_delete_elem = array_map_delete_elem,
275};
276
277static struct bpf_map_type_list array_type __read_mostly = {
278 .ops = &array_ops,
279 .type = BPF_MAP_TYPE_ARRAY,
280};
281
282static const struct bpf_map_ops percpu_array_ops = {
283 .map_alloc = array_map_alloc,
284 .map_free = array_map_free,
285 .map_get_next_key = array_map_get_next_key,
286 .map_lookup_elem = percpu_array_map_lookup_elem,
287 .map_update_elem = array_map_update_elem,
288 .map_delete_elem = array_map_delete_elem,
289};
290
291static struct bpf_map_type_list percpu_array_type __read_mostly = {
292 .ops = &percpu_array_ops,
293 .type = BPF_MAP_TYPE_PERCPU_ARRAY,
294};
295
296static int __init register_array_map(void)
297{
298 bpf_register_map_type(&array_type);
299 bpf_register_map_type(&percpu_array_type);
300 return 0;
301}
302late_initcall(register_array_map);
303
304static struct bpf_map *fd_array_map_alloc(union bpf_attr *attr)
305{
306 /* only file descriptors can be stored in this type of map */
307 if (attr->value_size != sizeof(u32))
308 return ERR_PTR(-EINVAL);
309 return array_map_alloc(attr);
310}
311
312static void fd_array_map_free(struct bpf_map *map)
313{
314 struct bpf_array *array = container_of(map, struct bpf_array, map);
315 int i;
316
317 synchronize_rcu();
318
319 /* make sure it's empty */
320 for (i = 0; i < array->map.max_entries; i++)
321 BUG_ON(array->ptrs[i] != NULL);
322 kvfree(array);
323}
324
325static void *fd_array_map_lookup_elem(struct bpf_map *map, void *key)
326{
327 return NULL;
328}
329
330/* only called from syscall */
331static int fd_array_map_update_elem(struct bpf_map *map, void *key,
332 void *value, u64 map_flags)
333{
334 struct bpf_array *array = container_of(map, struct bpf_array, map);
335 void *new_ptr, *old_ptr;
336 u32 index = *(u32 *)key, ufd;
337
338 if (map_flags != BPF_ANY)
339 return -EINVAL;
340
341 if (index >= array->map.max_entries)
342 return -E2BIG;
343
344 ufd = *(u32 *)value;
345 new_ptr = map->ops->map_fd_get_ptr(map, ufd);
346 if (IS_ERR(new_ptr))
347 return PTR_ERR(new_ptr);
348
349 old_ptr = xchg(array->ptrs + index, new_ptr);
350 if (old_ptr)
351 map->ops->map_fd_put_ptr(old_ptr);
352
353 return 0;
354}
355
356static int fd_array_map_delete_elem(struct bpf_map *map, void *key)
357{
358 struct bpf_array *array = container_of(map, struct bpf_array, map);
359 void *old_ptr;
360 u32 index = *(u32 *)key;
361
362 if (index >= array->map.max_entries)
363 return -E2BIG;
364
365 old_ptr = xchg(array->ptrs + index, NULL);
366 if (old_ptr) {
367 map->ops->map_fd_put_ptr(old_ptr);
368 return 0;
369 } else {
370 return -ENOENT;
371 }
372}
373
374static void *prog_fd_array_get_ptr(struct bpf_map *map, int fd)
375{
376 struct bpf_array *array = container_of(map, struct bpf_array, map);
377 struct bpf_prog *prog = bpf_prog_get(fd);
378 if (IS_ERR(prog))
379 return prog;
380
381 if (!bpf_prog_array_compatible(array, prog)) {
382 bpf_prog_put(prog);
383 return ERR_PTR(-EINVAL);
384 }
385 return prog;
386}
387
388static void prog_fd_array_put_ptr(void *ptr)
389{
390 struct bpf_prog *prog = ptr;
391
392 bpf_prog_put_rcu(prog);
393}
394
395/* decrement refcnt of all bpf_progs that are stored in this map */
396void bpf_fd_array_map_clear(struct bpf_map *map)
397{
398 struct bpf_array *array = container_of(map, struct bpf_array, map);
399 int i;
400
401 for (i = 0; i < array->map.max_entries; i++)
402 fd_array_map_delete_elem(map, &i);
403}
404
405static const struct bpf_map_ops prog_array_ops = {
406 .map_alloc = fd_array_map_alloc,
407 .map_free = fd_array_map_free,
408 .map_get_next_key = array_map_get_next_key,
409 .map_lookup_elem = fd_array_map_lookup_elem,
410 .map_update_elem = fd_array_map_update_elem,
411 .map_delete_elem = fd_array_map_delete_elem,
412 .map_fd_get_ptr = prog_fd_array_get_ptr,
413 .map_fd_put_ptr = prog_fd_array_put_ptr,
414};
415
416static struct bpf_map_type_list prog_array_type __read_mostly = {
417 .ops = &prog_array_ops,
418 .type = BPF_MAP_TYPE_PROG_ARRAY,
419};
420
421static int __init register_prog_array_map(void)
422{
423 bpf_register_map_type(&prog_array_type);
424 return 0;
425}
426late_initcall(register_prog_array_map);
427
428static void perf_event_array_map_free(struct bpf_map *map)
429{
430 bpf_fd_array_map_clear(map);
431 fd_array_map_free(map);
432}
433
434static void *perf_event_fd_array_get_ptr(struct bpf_map *map, int fd)
435{
436 struct perf_event *event;
437 const struct perf_event_attr *attr;
438 struct file *file;
439
440 file = perf_event_get(fd);
441 if (IS_ERR(file))
442 return file;
443
444 event = file->private_data;
445
446 attr = perf_event_attrs(event);
447 if (IS_ERR(attr))
448 goto err;
449
450 if (attr->inherit)
451 goto err;
452
453 if (attr->type == PERF_TYPE_RAW)
454 return file;
455
456 if (attr->type == PERF_TYPE_HARDWARE)
457 return file;
458
459 if (attr->type == PERF_TYPE_SOFTWARE &&
460 attr->config == PERF_COUNT_SW_BPF_OUTPUT)
461 return file;
462err:
463 fput(file);
464 return ERR_PTR(-EINVAL);
465}
466
467static void perf_event_fd_array_put_ptr(void *ptr)
468{
469 fput((struct file *)ptr);
470}
471
472static const struct bpf_map_ops perf_event_array_ops = {
473 .map_alloc = fd_array_map_alloc,
474 .map_free = perf_event_array_map_free,
475 .map_get_next_key = array_map_get_next_key,
476 .map_lookup_elem = fd_array_map_lookup_elem,
477 .map_update_elem = fd_array_map_update_elem,
478 .map_delete_elem = fd_array_map_delete_elem,
479 .map_fd_get_ptr = perf_event_fd_array_get_ptr,
480 .map_fd_put_ptr = perf_event_fd_array_put_ptr,
481};
482
483static struct bpf_map_type_list perf_event_array_type __read_mostly = {
484 .ops = &perf_event_array_ops,
485 .type = BPF_MAP_TYPE_PERF_EVENT_ARRAY,
486};
487
488static int __init register_perf_event_array_map(void)
489{
490 bpf_register_map_type(&perf_event_array_type);
491 return 0;
492}
493late_initcall(register_perf_event_array_map);
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
16#include "map_in_map.h"
17
18#define ARRAY_CREATE_FLAG_MASK \
19 (BPF_F_NUMA_NODE | BPF_F_MMAPABLE | BPF_F_ACCESS_MASK | \
20 BPF_F_PRESERVE_ELEMS | BPF_F_INNER_MAP)
21
22static void bpf_array_free_percpu(struct bpf_array *array)
23{
24 int i;
25
26 for (i = 0; i < array->map.max_entries; i++) {
27 free_percpu(array->pptrs[i]);
28 cond_resched();
29 }
30}
31
32static int bpf_array_alloc_percpu(struct bpf_array *array)
33{
34 void __percpu *ptr;
35 int i;
36
37 for (i = 0; i < array->map.max_entries; i++) {
38 ptr = bpf_map_alloc_percpu(&array->map, array->elem_size, 8,
39 GFP_USER | __GFP_NOWARN);
40 if (!ptr) {
41 bpf_array_free_percpu(array);
42 return -ENOMEM;
43 }
44 array->pptrs[i] = ptr;
45 cond_resched();
46 }
47
48 return 0;
49}
50
51/* Called from syscall */
52int array_map_alloc_check(union bpf_attr *attr)
53{
54 bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
55 int numa_node = bpf_map_attr_numa_node(attr);
56
57 /* check sanity of attributes */
58 if (attr->max_entries == 0 || attr->key_size != 4 ||
59 attr->value_size == 0 ||
60 attr->map_flags & ~ARRAY_CREATE_FLAG_MASK ||
61 !bpf_map_flags_access_ok(attr->map_flags) ||
62 (percpu && numa_node != NUMA_NO_NODE))
63 return -EINVAL;
64
65 if (attr->map_type != BPF_MAP_TYPE_ARRAY &&
66 attr->map_flags & (BPF_F_MMAPABLE | BPF_F_INNER_MAP))
67 return -EINVAL;
68
69 if (attr->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY &&
70 attr->map_flags & BPF_F_PRESERVE_ELEMS)
71 return -EINVAL;
72
73 /* avoid overflow on round_up(map->value_size) */
74 if (attr->value_size > INT_MAX)
75 return -E2BIG;
76
77 return 0;
78}
79
80static struct bpf_map *array_map_alloc(union bpf_attr *attr)
81{
82 bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
83 int numa_node = bpf_map_attr_numa_node(attr);
84 u32 elem_size, index_mask, max_entries;
85 bool bypass_spec_v1 = bpf_bypass_spec_v1();
86 u64 array_size, mask64;
87 struct bpf_array *array;
88
89 elem_size = round_up(attr->value_size, 8);
90
91 max_entries = attr->max_entries;
92
93 /* On 32 bit archs roundup_pow_of_two() with max_entries that has
94 * upper most bit set in u32 space is undefined behavior due to
95 * resulting 1U << 32, so do it manually here in u64 space.
96 */
97 mask64 = fls_long(max_entries - 1);
98 mask64 = 1ULL << mask64;
99 mask64 -= 1;
100
101 index_mask = mask64;
102 if (!bypass_spec_v1) {
103 /* round up array size to nearest power of 2,
104 * since cpu will speculate within index_mask limits
105 */
106 max_entries = index_mask + 1;
107 /* Check for overflows. */
108 if (max_entries < attr->max_entries)
109 return ERR_PTR(-E2BIG);
110 }
111
112 array_size = sizeof(*array);
113 if (percpu) {
114 array_size += (u64) max_entries * sizeof(void *);
115 } else {
116 /* rely on vmalloc() to return page-aligned memory and
117 * ensure array->value is exactly page-aligned
118 */
119 if (attr->map_flags & BPF_F_MMAPABLE) {
120 array_size = PAGE_ALIGN(array_size);
121 array_size += PAGE_ALIGN((u64) max_entries * elem_size);
122 } else {
123 array_size += (u64) max_entries * elem_size;
124 }
125 }
126
127 /* allocate all map elements and zero-initialize them */
128 if (attr->map_flags & BPF_F_MMAPABLE) {
129 void *data;
130
131 /* kmalloc'ed memory can't be mmap'ed, use explicit vmalloc */
132 data = bpf_map_area_mmapable_alloc(array_size, numa_node);
133 if (!data)
134 return ERR_PTR(-ENOMEM);
135 array = data + PAGE_ALIGN(sizeof(struct bpf_array))
136 - offsetof(struct bpf_array, value);
137 } else {
138 array = bpf_map_area_alloc(array_size, numa_node);
139 }
140 if (!array)
141 return ERR_PTR(-ENOMEM);
142 array->index_mask = index_mask;
143 array->map.bypass_spec_v1 = bypass_spec_v1;
144
145 /* copy mandatory map attributes */
146 bpf_map_init_from_attr(&array->map, attr);
147 array->elem_size = elem_size;
148
149 if (percpu && bpf_array_alloc_percpu(array)) {
150 bpf_map_area_free(array);
151 return ERR_PTR(-ENOMEM);
152 }
153
154 return &array->map;
155}
156
157static void *array_map_elem_ptr(struct bpf_array* array, u32 index)
158{
159 return array->value + (u64)array->elem_size * index;
160}
161
162/* Called from syscall or from eBPF program */
163static void *array_map_lookup_elem(struct bpf_map *map, void *key)
164{
165 struct bpf_array *array = container_of(map, struct bpf_array, map);
166 u32 index = *(u32 *)key;
167
168 if (unlikely(index >= array->map.max_entries))
169 return NULL;
170
171 return array->value + (u64)array->elem_size * (index & array->index_mask);
172}
173
174static int array_map_direct_value_addr(const struct bpf_map *map, u64 *imm,
175 u32 off)
176{
177 struct bpf_array *array = container_of(map, struct bpf_array, map);
178
179 if (map->max_entries != 1)
180 return -ENOTSUPP;
181 if (off >= map->value_size)
182 return -EINVAL;
183
184 *imm = (unsigned long)array->value;
185 return 0;
186}
187
188static int array_map_direct_value_meta(const struct bpf_map *map, u64 imm,
189 u32 *off)
190{
191 struct bpf_array *array = container_of(map, struct bpf_array, map);
192 u64 base = (unsigned long)array->value;
193 u64 range = array->elem_size;
194
195 if (map->max_entries != 1)
196 return -ENOTSUPP;
197 if (imm < base || imm >= base + range)
198 return -ENOENT;
199
200 *off = imm - base;
201 return 0;
202}
203
204/* emit BPF instructions equivalent to C code of array_map_lookup_elem() */
205static int array_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
206{
207 struct bpf_array *array = container_of(map, struct bpf_array, map);
208 struct bpf_insn *insn = insn_buf;
209 u32 elem_size = array->elem_size;
210 const int ret = BPF_REG_0;
211 const int map_ptr = BPF_REG_1;
212 const int index = BPF_REG_2;
213
214 if (map->map_flags & BPF_F_INNER_MAP)
215 return -EOPNOTSUPP;
216
217 *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
218 *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
219 if (!map->bypass_spec_v1) {
220 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 4);
221 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
222 } else {
223 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 3);
224 }
225
226 if (is_power_of_2(elem_size)) {
227 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
228 } else {
229 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
230 }
231 *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
232 *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
233 *insn++ = BPF_MOV64_IMM(ret, 0);
234 return insn - insn_buf;
235}
236
237/* Called from eBPF program */
238static void *percpu_array_map_lookup_elem(struct bpf_map *map, void *key)
239{
240 struct bpf_array *array = container_of(map, struct bpf_array, map);
241 u32 index = *(u32 *)key;
242
243 if (unlikely(index >= array->map.max_entries))
244 return NULL;
245
246 return this_cpu_ptr(array->pptrs[index & array->index_mask]);
247}
248
249static void *percpu_array_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
250{
251 struct bpf_array *array = container_of(map, struct bpf_array, map);
252 u32 index = *(u32 *)key;
253
254 if (cpu >= nr_cpu_ids)
255 return NULL;
256
257 if (unlikely(index >= array->map.max_entries))
258 return NULL;
259
260 return per_cpu_ptr(array->pptrs[index & array->index_mask], cpu);
261}
262
263int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value)
264{
265 struct bpf_array *array = container_of(map, struct bpf_array, map);
266 u32 index = *(u32 *)key;
267 void __percpu *pptr;
268 int cpu, off = 0;
269 u32 size;
270
271 if (unlikely(index >= array->map.max_entries))
272 return -ENOENT;
273
274 /* per_cpu areas are zero-filled and bpf programs can only
275 * access 'value_size' of them, so copying rounded areas
276 * will not leak any kernel data
277 */
278 size = array->elem_size;
279 rcu_read_lock();
280 pptr = array->pptrs[index & array->index_mask];
281 for_each_possible_cpu(cpu) {
282 copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
283 check_and_init_map_value(map, value + off);
284 off += size;
285 }
286 rcu_read_unlock();
287 return 0;
288}
289
290/* Called from syscall */
291static int array_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
292{
293 struct bpf_array *array = container_of(map, struct bpf_array, map);
294 u32 index = key ? *(u32 *)key : U32_MAX;
295 u32 *next = (u32 *)next_key;
296
297 if (index >= array->map.max_entries) {
298 *next = 0;
299 return 0;
300 }
301
302 if (index == array->map.max_entries - 1)
303 return -ENOENT;
304
305 *next = index + 1;
306 return 0;
307}
308
309/* Called from syscall or from eBPF program */
310static int array_map_update_elem(struct bpf_map *map, void *key, void *value,
311 u64 map_flags)
312{
313 struct bpf_array *array = container_of(map, struct bpf_array, map);
314 u32 index = *(u32 *)key;
315 char *val;
316
317 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
318 /* unknown flags */
319 return -EINVAL;
320
321 if (unlikely(index >= array->map.max_entries))
322 /* all elements were pre-allocated, cannot insert a new one */
323 return -E2BIG;
324
325 if (unlikely(map_flags & BPF_NOEXIST))
326 /* all elements already exist */
327 return -EEXIST;
328
329 if (unlikely((map_flags & BPF_F_LOCK) &&
330 !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
331 return -EINVAL;
332
333 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
334 val = this_cpu_ptr(array->pptrs[index & array->index_mask]);
335 copy_map_value(map, val, value);
336 bpf_obj_free_fields(array->map.record, val);
337 } else {
338 val = array->value +
339 (u64)array->elem_size * (index & array->index_mask);
340 if (map_flags & BPF_F_LOCK)
341 copy_map_value_locked(map, val, value, false);
342 else
343 copy_map_value(map, val, value);
344 bpf_obj_free_fields(array->map.record, val);
345 }
346 return 0;
347}
348
349int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
350 u64 map_flags)
351{
352 struct bpf_array *array = container_of(map, struct bpf_array, map);
353 u32 index = *(u32 *)key;
354 void __percpu *pptr;
355 int cpu, off = 0;
356 u32 size;
357
358 if (unlikely(map_flags > BPF_EXIST))
359 /* unknown flags */
360 return -EINVAL;
361
362 if (unlikely(index >= array->map.max_entries))
363 /* all elements were pre-allocated, cannot insert a new one */
364 return -E2BIG;
365
366 if (unlikely(map_flags == BPF_NOEXIST))
367 /* all elements already exist */
368 return -EEXIST;
369
370 /* the user space will provide round_up(value_size, 8) bytes that
371 * will be copied into per-cpu area. bpf programs can only access
372 * value_size of it. During lookup the same extra bytes will be
373 * returned or zeros which were zero-filled by percpu_alloc,
374 * so no kernel data leaks possible
375 */
376 size = array->elem_size;
377 rcu_read_lock();
378 pptr = array->pptrs[index & array->index_mask];
379 for_each_possible_cpu(cpu) {
380 copy_map_value_long(map, per_cpu_ptr(pptr, cpu), value + off);
381 bpf_obj_free_fields(array->map.record, per_cpu_ptr(pptr, cpu));
382 off += size;
383 }
384 rcu_read_unlock();
385 return 0;
386}
387
388/* Called from syscall or from eBPF program */
389static int array_map_delete_elem(struct bpf_map *map, void *key)
390{
391 return -EINVAL;
392}
393
394static void *array_map_vmalloc_addr(struct bpf_array *array)
395{
396 return (void *)round_down((unsigned long)array, PAGE_SIZE);
397}
398
399static void array_map_free_timers(struct bpf_map *map)
400{
401 struct bpf_array *array = container_of(map, struct bpf_array, map);
402 int i;
403
404 /* We don't reset or free fields other than timer on uref dropping to zero. */
405 if (!btf_record_has_field(map->record, BPF_TIMER))
406 return;
407
408 for (i = 0; i < array->map.max_entries; i++)
409 bpf_obj_free_timer(map->record, array_map_elem_ptr(array, i));
410}
411
412/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
413static void array_map_free(struct bpf_map *map)
414{
415 struct bpf_array *array = container_of(map, struct bpf_array, map);
416 int i;
417
418 if (!IS_ERR_OR_NULL(map->record)) {
419 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
420 for (i = 0; i < array->map.max_entries; i++) {
421 void __percpu *pptr = array->pptrs[i & array->index_mask];
422 int cpu;
423
424 for_each_possible_cpu(cpu) {
425 bpf_obj_free_fields(map->record, per_cpu_ptr(pptr, cpu));
426 cond_resched();
427 }
428 }
429 } else {
430 for (i = 0; i < array->map.max_entries; i++)
431 bpf_obj_free_fields(map->record, array_map_elem_ptr(array, i));
432 }
433 }
434
435 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
436 bpf_array_free_percpu(array);
437
438 if (array->map.map_flags & BPF_F_MMAPABLE)
439 bpf_map_area_free(array_map_vmalloc_addr(array));
440 else
441 bpf_map_area_free(array);
442}
443
444static void array_map_seq_show_elem(struct bpf_map *map, void *key,
445 struct seq_file *m)
446{
447 void *value;
448
449 rcu_read_lock();
450
451 value = array_map_lookup_elem(map, key);
452 if (!value) {
453 rcu_read_unlock();
454 return;
455 }
456
457 if (map->btf_key_type_id)
458 seq_printf(m, "%u: ", *(u32 *)key);
459 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
460 seq_puts(m, "\n");
461
462 rcu_read_unlock();
463}
464
465static void percpu_array_map_seq_show_elem(struct bpf_map *map, void *key,
466 struct seq_file *m)
467{
468 struct bpf_array *array = container_of(map, struct bpf_array, map);
469 u32 index = *(u32 *)key;
470 void __percpu *pptr;
471 int cpu;
472
473 rcu_read_lock();
474
475 seq_printf(m, "%u: {\n", *(u32 *)key);
476 pptr = array->pptrs[index & array->index_mask];
477 for_each_possible_cpu(cpu) {
478 seq_printf(m, "\tcpu%d: ", cpu);
479 btf_type_seq_show(map->btf, map->btf_value_type_id,
480 per_cpu_ptr(pptr, cpu), m);
481 seq_puts(m, "\n");
482 }
483 seq_puts(m, "}\n");
484
485 rcu_read_unlock();
486}
487
488static int array_map_check_btf(const struct bpf_map *map,
489 const struct btf *btf,
490 const struct btf_type *key_type,
491 const struct btf_type *value_type)
492{
493 u32 int_data;
494
495 /* One exception for keyless BTF: .bss/.data/.rodata map */
496 if (btf_type_is_void(key_type)) {
497 if (map->map_type != BPF_MAP_TYPE_ARRAY ||
498 map->max_entries != 1)
499 return -EINVAL;
500
501 if (BTF_INFO_KIND(value_type->info) != BTF_KIND_DATASEC)
502 return -EINVAL;
503
504 return 0;
505 }
506
507 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
508 return -EINVAL;
509
510 int_data = *(u32 *)(key_type + 1);
511 /* bpf array can only take a u32 key. This check makes sure
512 * that the btf matches the attr used during map_create.
513 */
514 if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
515 return -EINVAL;
516
517 return 0;
518}
519
520static int array_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
521{
522 struct bpf_array *array = container_of(map, struct bpf_array, map);
523 pgoff_t pgoff = PAGE_ALIGN(sizeof(*array)) >> PAGE_SHIFT;
524
525 if (!(map->map_flags & BPF_F_MMAPABLE))
526 return -EINVAL;
527
528 if (vma->vm_pgoff * PAGE_SIZE + (vma->vm_end - vma->vm_start) >
529 PAGE_ALIGN((u64)array->map.max_entries * array->elem_size))
530 return -EINVAL;
531
532 return remap_vmalloc_range(vma, array_map_vmalloc_addr(array),
533 vma->vm_pgoff + pgoff);
534}
535
536static bool array_map_meta_equal(const struct bpf_map *meta0,
537 const struct bpf_map *meta1)
538{
539 if (!bpf_map_meta_equal(meta0, meta1))
540 return false;
541 return meta0->map_flags & BPF_F_INNER_MAP ? true :
542 meta0->max_entries == meta1->max_entries;
543}
544
545struct bpf_iter_seq_array_map_info {
546 struct bpf_map *map;
547 void *percpu_value_buf;
548 u32 index;
549};
550
551static void *bpf_array_map_seq_start(struct seq_file *seq, loff_t *pos)
552{
553 struct bpf_iter_seq_array_map_info *info = seq->private;
554 struct bpf_map *map = info->map;
555 struct bpf_array *array;
556 u32 index;
557
558 if (info->index >= map->max_entries)
559 return NULL;
560
561 if (*pos == 0)
562 ++*pos;
563 array = container_of(map, struct bpf_array, map);
564 index = info->index & array->index_mask;
565 if (info->percpu_value_buf)
566 return array->pptrs[index];
567 return array_map_elem_ptr(array, index);
568}
569
570static void *bpf_array_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
571{
572 struct bpf_iter_seq_array_map_info *info = seq->private;
573 struct bpf_map *map = info->map;
574 struct bpf_array *array;
575 u32 index;
576
577 ++*pos;
578 ++info->index;
579 if (info->index >= map->max_entries)
580 return NULL;
581
582 array = container_of(map, struct bpf_array, map);
583 index = info->index & array->index_mask;
584 if (info->percpu_value_buf)
585 return array->pptrs[index];
586 return array_map_elem_ptr(array, index);
587}
588
589static int __bpf_array_map_seq_show(struct seq_file *seq, void *v)
590{
591 struct bpf_iter_seq_array_map_info *info = seq->private;
592 struct bpf_iter__bpf_map_elem ctx = {};
593 struct bpf_map *map = info->map;
594 struct bpf_array *array = container_of(map, struct bpf_array, map);
595 struct bpf_iter_meta meta;
596 struct bpf_prog *prog;
597 int off = 0, cpu = 0;
598 void __percpu **pptr;
599 u32 size;
600
601 meta.seq = seq;
602 prog = bpf_iter_get_info(&meta, v == NULL);
603 if (!prog)
604 return 0;
605
606 ctx.meta = &meta;
607 ctx.map = info->map;
608 if (v) {
609 ctx.key = &info->index;
610
611 if (!info->percpu_value_buf) {
612 ctx.value = v;
613 } else {
614 pptr = v;
615 size = array->elem_size;
616 for_each_possible_cpu(cpu) {
617 copy_map_value_long(map, info->percpu_value_buf + off,
618 per_cpu_ptr(pptr, cpu));
619 check_and_init_map_value(map, info->percpu_value_buf + off);
620 off += size;
621 }
622 ctx.value = info->percpu_value_buf;
623 }
624 }
625
626 return bpf_iter_run_prog(prog, &ctx);
627}
628
629static int bpf_array_map_seq_show(struct seq_file *seq, void *v)
630{
631 return __bpf_array_map_seq_show(seq, v);
632}
633
634static void bpf_array_map_seq_stop(struct seq_file *seq, void *v)
635{
636 if (!v)
637 (void)__bpf_array_map_seq_show(seq, NULL);
638}
639
640static int bpf_iter_init_array_map(void *priv_data,
641 struct bpf_iter_aux_info *aux)
642{
643 struct bpf_iter_seq_array_map_info *seq_info = priv_data;
644 struct bpf_map *map = aux->map;
645 struct bpf_array *array = container_of(map, struct bpf_array, map);
646 void *value_buf;
647 u32 buf_size;
648
649 if (map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
650 buf_size = array->elem_size * num_possible_cpus();
651 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
652 if (!value_buf)
653 return -ENOMEM;
654
655 seq_info->percpu_value_buf = value_buf;
656 }
657
658 /* bpf_iter_attach_map() acquires a map uref, and the uref may be
659 * released before or in the middle of iterating map elements, so
660 * acquire an extra map uref for iterator.
661 */
662 bpf_map_inc_with_uref(map);
663 seq_info->map = map;
664 return 0;
665}
666
667static void bpf_iter_fini_array_map(void *priv_data)
668{
669 struct bpf_iter_seq_array_map_info *seq_info = priv_data;
670
671 bpf_map_put_with_uref(seq_info->map);
672 kfree(seq_info->percpu_value_buf);
673}
674
675static const struct seq_operations bpf_array_map_seq_ops = {
676 .start = bpf_array_map_seq_start,
677 .next = bpf_array_map_seq_next,
678 .stop = bpf_array_map_seq_stop,
679 .show = bpf_array_map_seq_show,
680};
681
682static const struct bpf_iter_seq_info iter_seq_info = {
683 .seq_ops = &bpf_array_map_seq_ops,
684 .init_seq_private = bpf_iter_init_array_map,
685 .fini_seq_private = bpf_iter_fini_array_map,
686 .seq_priv_size = sizeof(struct bpf_iter_seq_array_map_info),
687};
688
689static int bpf_for_each_array_elem(struct bpf_map *map, bpf_callback_t callback_fn,
690 void *callback_ctx, u64 flags)
691{
692 u32 i, key, num_elems = 0;
693 struct bpf_array *array;
694 bool is_percpu;
695 u64 ret = 0;
696 void *val;
697
698 if (flags != 0)
699 return -EINVAL;
700
701 is_percpu = map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
702 array = container_of(map, struct bpf_array, map);
703 if (is_percpu)
704 migrate_disable();
705 for (i = 0; i < map->max_entries; i++) {
706 if (is_percpu)
707 val = this_cpu_ptr(array->pptrs[i]);
708 else
709 val = array_map_elem_ptr(array, i);
710 num_elems++;
711 key = i;
712 ret = callback_fn((u64)(long)map, (u64)(long)&key,
713 (u64)(long)val, (u64)(long)callback_ctx, 0);
714 /* return value: 0 - continue, 1 - stop and return */
715 if (ret)
716 break;
717 }
718
719 if (is_percpu)
720 migrate_enable();
721 return num_elems;
722}
723
724BTF_ID_LIST_SINGLE(array_map_btf_ids, struct, bpf_array)
725const struct bpf_map_ops array_map_ops = {
726 .map_meta_equal = array_map_meta_equal,
727 .map_alloc_check = array_map_alloc_check,
728 .map_alloc = array_map_alloc,
729 .map_free = array_map_free,
730 .map_get_next_key = array_map_get_next_key,
731 .map_release_uref = array_map_free_timers,
732 .map_lookup_elem = array_map_lookup_elem,
733 .map_update_elem = array_map_update_elem,
734 .map_delete_elem = array_map_delete_elem,
735 .map_gen_lookup = array_map_gen_lookup,
736 .map_direct_value_addr = array_map_direct_value_addr,
737 .map_direct_value_meta = array_map_direct_value_meta,
738 .map_mmap = array_map_mmap,
739 .map_seq_show_elem = array_map_seq_show_elem,
740 .map_check_btf = array_map_check_btf,
741 .map_lookup_batch = generic_map_lookup_batch,
742 .map_update_batch = generic_map_update_batch,
743 .map_set_for_each_callback_args = map_set_for_each_callback_args,
744 .map_for_each_callback = bpf_for_each_array_elem,
745 .map_btf_id = &array_map_btf_ids[0],
746 .iter_seq_info = &iter_seq_info,
747};
748
749const struct bpf_map_ops percpu_array_map_ops = {
750 .map_meta_equal = bpf_map_meta_equal,
751 .map_alloc_check = array_map_alloc_check,
752 .map_alloc = array_map_alloc,
753 .map_free = array_map_free,
754 .map_get_next_key = array_map_get_next_key,
755 .map_lookup_elem = percpu_array_map_lookup_elem,
756 .map_update_elem = array_map_update_elem,
757 .map_delete_elem = array_map_delete_elem,
758 .map_lookup_percpu_elem = percpu_array_map_lookup_percpu_elem,
759 .map_seq_show_elem = percpu_array_map_seq_show_elem,
760 .map_check_btf = array_map_check_btf,
761 .map_lookup_batch = generic_map_lookup_batch,
762 .map_update_batch = generic_map_update_batch,
763 .map_set_for_each_callback_args = map_set_for_each_callback_args,
764 .map_for_each_callback = bpf_for_each_array_elem,
765 .map_btf_id = &array_map_btf_ids[0],
766 .iter_seq_info = &iter_seq_info,
767};
768
769static int fd_array_map_alloc_check(union bpf_attr *attr)
770{
771 /* only file descriptors can be stored in this type of map */
772 if (attr->value_size != sizeof(u32))
773 return -EINVAL;
774 /* Program read-only/write-only not supported for special maps yet. */
775 if (attr->map_flags & (BPF_F_RDONLY_PROG | BPF_F_WRONLY_PROG))
776 return -EINVAL;
777 return array_map_alloc_check(attr);
778}
779
780static void fd_array_map_free(struct bpf_map *map)
781{
782 struct bpf_array *array = container_of(map, struct bpf_array, map);
783 int i;
784
785 /* make sure it's empty */
786 for (i = 0; i < array->map.max_entries; i++)
787 BUG_ON(array->ptrs[i] != NULL);
788
789 bpf_map_area_free(array);
790}
791
792static void *fd_array_map_lookup_elem(struct bpf_map *map, void *key)
793{
794 return ERR_PTR(-EOPNOTSUPP);
795}
796
797/* only called from syscall */
798int bpf_fd_array_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
799{
800 void **elem, *ptr;
801 int ret = 0;
802
803 if (!map->ops->map_fd_sys_lookup_elem)
804 return -ENOTSUPP;
805
806 rcu_read_lock();
807 elem = array_map_lookup_elem(map, key);
808 if (elem && (ptr = READ_ONCE(*elem)))
809 *value = map->ops->map_fd_sys_lookup_elem(ptr);
810 else
811 ret = -ENOENT;
812 rcu_read_unlock();
813
814 return ret;
815}
816
817/* only called from syscall */
818int bpf_fd_array_map_update_elem(struct bpf_map *map, struct file *map_file,
819 void *key, void *value, u64 map_flags)
820{
821 struct bpf_array *array = container_of(map, struct bpf_array, map);
822 void *new_ptr, *old_ptr;
823 u32 index = *(u32 *)key, ufd;
824
825 if (map_flags != BPF_ANY)
826 return -EINVAL;
827
828 if (index >= array->map.max_entries)
829 return -E2BIG;
830
831 ufd = *(u32 *)value;
832 new_ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
833 if (IS_ERR(new_ptr))
834 return PTR_ERR(new_ptr);
835
836 if (map->ops->map_poke_run) {
837 mutex_lock(&array->aux->poke_mutex);
838 old_ptr = xchg(array->ptrs + index, new_ptr);
839 map->ops->map_poke_run(map, index, old_ptr, new_ptr);
840 mutex_unlock(&array->aux->poke_mutex);
841 } else {
842 old_ptr = xchg(array->ptrs + index, new_ptr);
843 }
844
845 if (old_ptr)
846 map->ops->map_fd_put_ptr(old_ptr);
847 return 0;
848}
849
850static int fd_array_map_delete_elem(struct bpf_map *map, void *key)
851{
852 struct bpf_array *array = container_of(map, struct bpf_array, map);
853 void *old_ptr;
854 u32 index = *(u32 *)key;
855
856 if (index >= array->map.max_entries)
857 return -E2BIG;
858
859 if (map->ops->map_poke_run) {
860 mutex_lock(&array->aux->poke_mutex);
861 old_ptr = xchg(array->ptrs + index, NULL);
862 map->ops->map_poke_run(map, index, old_ptr, NULL);
863 mutex_unlock(&array->aux->poke_mutex);
864 } else {
865 old_ptr = xchg(array->ptrs + index, NULL);
866 }
867
868 if (old_ptr) {
869 map->ops->map_fd_put_ptr(old_ptr);
870 return 0;
871 } else {
872 return -ENOENT;
873 }
874}
875
876static void *prog_fd_array_get_ptr(struct bpf_map *map,
877 struct file *map_file, int fd)
878{
879 struct bpf_prog *prog = bpf_prog_get(fd);
880
881 if (IS_ERR(prog))
882 return prog;
883
884 if (!bpf_prog_map_compatible(map, prog)) {
885 bpf_prog_put(prog);
886 return ERR_PTR(-EINVAL);
887 }
888
889 return prog;
890}
891
892static void prog_fd_array_put_ptr(void *ptr)
893{
894 bpf_prog_put(ptr);
895}
896
897static u32 prog_fd_array_sys_lookup_elem(void *ptr)
898{
899 return ((struct bpf_prog *)ptr)->aux->id;
900}
901
902/* decrement refcnt of all bpf_progs that are stored in this map */
903static void bpf_fd_array_map_clear(struct bpf_map *map)
904{
905 struct bpf_array *array = container_of(map, struct bpf_array, map);
906 int i;
907
908 for (i = 0; i < array->map.max_entries; i++)
909 fd_array_map_delete_elem(map, &i);
910}
911
912static void prog_array_map_seq_show_elem(struct bpf_map *map, void *key,
913 struct seq_file *m)
914{
915 void **elem, *ptr;
916 u32 prog_id;
917
918 rcu_read_lock();
919
920 elem = array_map_lookup_elem(map, key);
921 if (elem) {
922 ptr = READ_ONCE(*elem);
923 if (ptr) {
924 seq_printf(m, "%u: ", *(u32 *)key);
925 prog_id = prog_fd_array_sys_lookup_elem(ptr);
926 btf_type_seq_show(map->btf, map->btf_value_type_id,
927 &prog_id, m);
928 seq_puts(m, "\n");
929 }
930 }
931
932 rcu_read_unlock();
933}
934
935struct prog_poke_elem {
936 struct list_head list;
937 struct bpf_prog_aux *aux;
938};
939
940static int prog_array_map_poke_track(struct bpf_map *map,
941 struct bpf_prog_aux *prog_aux)
942{
943 struct prog_poke_elem *elem;
944 struct bpf_array_aux *aux;
945 int ret = 0;
946
947 aux = container_of(map, struct bpf_array, map)->aux;
948 mutex_lock(&aux->poke_mutex);
949 list_for_each_entry(elem, &aux->poke_progs, list) {
950 if (elem->aux == prog_aux)
951 goto out;
952 }
953
954 elem = kmalloc(sizeof(*elem), GFP_KERNEL);
955 if (!elem) {
956 ret = -ENOMEM;
957 goto out;
958 }
959
960 INIT_LIST_HEAD(&elem->list);
961 /* We must track the program's aux info at this point in time
962 * since the program pointer itself may not be stable yet, see
963 * also comment in prog_array_map_poke_run().
964 */
965 elem->aux = prog_aux;
966
967 list_add_tail(&elem->list, &aux->poke_progs);
968out:
969 mutex_unlock(&aux->poke_mutex);
970 return ret;
971}
972
973static void prog_array_map_poke_untrack(struct bpf_map *map,
974 struct bpf_prog_aux *prog_aux)
975{
976 struct prog_poke_elem *elem, *tmp;
977 struct bpf_array_aux *aux;
978
979 aux = container_of(map, struct bpf_array, map)->aux;
980 mutex_lock(&aux->poke_mutex);
981 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
982 if (elem->aux == prog_aux) {
983 list_del_init(&elem->list);
984 kfree(elem);
985 break;
986 }
987 }
988 mutex_unlock(&aux->poke_mutex);
989}
990
991static void prog_array_map_poke_run(struct bpf_map *map, u32 key,
992 struct bpf_prog *old,
993 struct bpf_prog *new)
994{
995 u8 *old_addr, *new_addr, *old_bypass_addr;
996 struct prog_poke_elem *elem;
997 struct bpf_array_aux *aux;
998
999 aux = container_of(map, struct bpf_array, map)->aux;
1000 WARN_ON_ONCE(!mutex_is_locked(&aux->poke_mutex));
1001
1002 list_for_each_entry(elem, &aux->poke_progs, list) {
1003 struct bpf_jit_poke_descriptor *poke;
1004 int i, ret;
1005
1006 for (i = 0; i < elem->aux->size_poke_tab; i++) {
1007 poke = &elem->aux->poke_tab[i];
1008
1009 /* Few things to be aware of:
1010 *
1011 * 1) We can only ever access aux in this context, but
1012 * not aux->prog since it might not be stable yet and
1013 * there could be danger of use after free otherwise.
1014 * 2) Initially when we start tracking aux, the program
1015 * is not JITed yet and also does not have a kallsyms
1016 * entry. We skip these as poke->tailcall_target_stable
1017 * is not active yet. The JIT will do the final fixup
1018 * before setting it stable. The various
1019 * poke->tailcall_target_stable are successively
1020 * activated, so tail call updates can arrive from here
1021 * while JIT is still finishing its final fixup for
1022 * non-activated poke entries.
1023 * 3) On program teardown, the program's kallsym entry gets
1024 * removed out of RCU callback, but we can only untrack
1025 * from sleepable context, therefore bpf_arch_text_poke()
1026 * might not see that this is in BPF text section and
1027 * bails out with -EINVAL. As these are unreachable since
1028 * RCU grace period already passed, we simply skip them.
1029 * 4) Also programs reaching refcount of zero while patching
1030 * is in progress is okay since we're protected under
1031 * poke_mutex and untrack the programs before the JIT
1032 * buffer is freed. When we're still in the middle of
1033 * patching and suddenly kallsyms entry of the program
1034 * gets evicted, we just skip the rest which is fine due
1035 * to point 3).
1036 * 5) Any other error happening below from bpf_arch_text_poke()
1037 * is a unexpected bug.
1038 */
1039 if (!READ_ONCE(poke->tailcall_target_stable))
1040 continue;
1041 if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
1042 continue;
1043 if (poke->tail_call.map != map ||
1044 poke->tail_call.key != key)
1045 continue;
1046
1047 old_bypass_addr = old ? NULL : poke->bypass_addr;
1048 old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
1049 new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
1050
1051 if (new) {
1052 ret = bpf_arch_text_poke(poke->tailcall_target,
1053 BPF_MOD_JUMP,
1054 old_addr, new_addr);
1055 BUG_ON(ret < 0 && ret != -EINVAL);
1056 if (!old) {
1057 ret = bpf_arch_text_poke(poke->tailcall_bypass,
1058 BPF_MOD_JUMP,
1059 poke->bypass_addr,
1060 NULL);
1061 BUG_ON(ret < 0 && ret != -EINVAL);
1062 }
1063 } else {
1064 ret = bpf_arch_text_poke(poke->tailcall_bypass,
1065 BPF_MOD_JUMP,
1066 old_bypass_addr,
1067 poke->bypass_addr);
1068 BUG_ON(ret < 0 && ret != -EINVAL);
1069 /* let other CPUs finish the execution of program
1070 * so that it will not possible to expose them
1071 * to invalid nop, stack unwind, nop state
1072 */
1073 if (!ret)
1074 synchronize_rcu();
1075 ret = bpf_arch_text_poke(poke->tailcall_target,
1076 BPF_MOD_JUMP,
1077 old_addr, NULL);
1078 BUG_ON(ret < 0 && ret != -EINVAL);
1079 }
1080 }
1081 }
1082}
1083
1084static void prog_array_map_clear_deferred(struct work_struct *work)
1085{
1086 struct bpf_map *map = container_of(work, struct bpf_array_aux,
1087 work)->map;
1088 bpf_fd_array_map_clear(map);
1089 bpf_map_put(map);
1090}
1091
1092static void prog_array_map_clear(struct bpf_map *map)
1093{
1094 struct bpf_array_aux *aux = container_of(map, struct bpf_array,
1095 map)->aux;
1096 bpf_map_inc(map);
1097 schedule_work(&aux->work);
1098}
1099
1100static struct bpf_map *prog_array_map_alloc(union bpf_attr *attr)
1101{
1102 struct bpf_array_aux *aux;
1103 struct bpf_map *map;
1104
1105 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT);
1106 if (!aux)
1107 return ERR_PTR(-ENOMEM);
1108
1109 INIT_WORK(&aux->work, prog_array_map_clear_deferred);
1110 INIT_LIST_HEAD(&aux->poke_progs);
1111 mutex_init(&aux->poke_mutex);
1112
1113 map = array_map_alloc(attr);
1114 if (IS_ERR(map)) {
1115 kfree(aux);
1116 return map;
1117 }
1118
1119 container_of(map, struct bpf_array, map)->aux = aux;
1120 aux->map = map;
1121
1122 return map;
1123}
1124
1125static void prog_array_map_free(struct bpf_map *map)
1126{
1127 struct prog_poke_elem *elem, *tmp;
1128 struct bpf_array_aux *aux;
1129
1130 aux = container_of(map, struct bpf_array, map)->aux;
1131 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
1132 list_del_init(&elem->list);
1133 kfree(elem);
1134 }
1135 kfree(aux);
1136 fd_array_map_free(map);
1137}
1138
1139/* prog_array->aux->{type,jited} is a runtime binding.
1140 * Doing static check alone in the verifier is not enough.
1141 * Thus, prog_array_map cannot be used as an inner_map
1142 * and map_meta_equal is not implemented.
1143 */
1144const struct bpf_map_ops prog_array_map_ops = {
1145 .map_alloc_check = fd_array_map_alloc_check,
1146 .map_alloc = prog_array_map_alloc,
1147 .map_free = prog_array_map_free,
1148 .map_poke_track = prog_array_map_poke_track,
1149 .map_poke_untrack = prog_array_map_poke_untrack,
1150 .map_poke_run = prog_array_map_poke_run,
1151 .map_get_next_key = array_map_get_next_key,
1152 .map_lookup_elem = fd_array_map_lookup_elem,
1153 .map_delete_elem = fd_array_map_delete_elem,
1154 .map_fd_get_ptr = prog_fd_array_get_ptr,
1155 .map_fd_put_ptr = prog_fd_array_put_ptr,
1156 .map_fd_sys_lookup_elem = prog_fd_array_sys_lookup_elem,
1157 .map_release_uref = prog_array_map_clear,
1158 .map_seq_show_elem = prog_array_map_seq_show_elem,
1159 .map_btf_id = &array_map_btf_ids[0],
1160};
1161
1162static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
1163 struct file *map_file)
1164{
1165 struct bpf_event_entry *ee;
1166
1167 ee = kzalloc(sizeof(*ee), GFP_ATOMIC);
1168 if (ee) {
1169 ee->event = perf_file->private_data;
1170 ee->perf_file = perf_file;
1171 ee->map_file = map_file;
1172 }
1173
1174 return ee;
1175}
1176
1177static void __bpf_event_entry_free(struct rcu_head *rcu)
1178{
1179 struct bpf_event_entry *ee;
1180
1181 ee = container_of(rcu, struct bpf_event_entry, rcu);
1182 fput(ee->perf_file);
1183 kfree(ee);
1184}
1185
1186static void bpf_event_entry_free_rcu(struct bpf_event_entry *ee)
1187{
1188 call_rcu(&ee->rcu, __bpf_event_entry_free);
1189}
1190
1191static void *perf_event_fd_array_get_ptr(struct bpf_map *map,
1192 struct file *map_file, int fd)
1193{
1194 struct bpf_event_entry *ee;
1195 struct perf_event *event;
1196 struct file *perf_file;
1197 u64 value;
1198
1199 perf_file = perf_event_get(fd);
1200 if (IS_ERR(perf_file))
1201 return perf_file;
1202
1203 ee = ERR_PTR(-EOPNOTSUPP);
1204 event = perf_file->private_data;
1205 if (perf_event_read_local(event, &value, NULL, NULL) == -EOPNOTSUPP)
1206 goto err_out;
1207
1208 ee = bpf_event_entry_gen(perf_file, map_file);
1209 if (ee)
1210 return ee;
1211 ee = ERR_PTR(-ENOMEM);
1212err_out:
1213 fput(perf_file);
1214 return ee;
1215}
1216
1217static void perf_event_fd_array_put_ptr(void *ptr)
1218{
1219 bpf_event_entry_free_rcu(ptr);
1220}
1221
1222static void perf_event_fd_array_release(struct bpf_map *map,
1223 struct file *map_file)
1224{
1225 struct bpf_array *array = container_of(map, struct bpf_array, map);
1226 struct bpf_event_entry *ee;
1227 int i;
1228
1229 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1230 return;
1231
1232 rcu_read_lock();
1233 for (i = 0; i < array->map.max_entries; i++) {
1234 ee = READ_ONCE(array->ptrs[i]);
1235 if (ee && ee->map_file == map_file)
1236 fd_array_map_delete_elem(map, &i);
1237 }
1238 rcu_read_unlock();
1239}
1240
1241static void perf_event_fd_array_map_free(struct bpf_map *map)
1242{
1243 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1244 bpf_fd_array_map_clear(map);
1245 fd_array_map_free(map);
1246}
1247
1248const struct bpf_map_ops perf_event_array_map_ops = {
1249 .map_meta_equal = bpf_map_meta_equal,
1250 .map_alloc_check = fd_array_map_alloc_check,
1251 .map_alloc = array_map_alloc,
1252 .map_free = perf_event_fd_array_map_free,
1253 .map_get_next_key = array_map_get_next_key,
1254 .map_lookup_elem = fd_array_map_lookup_elem,
1255 .map_delete_elem = fd_array_map_delete_elem,
1256 .map_fd_get_ptr = perf_event_fd_array_get_ptr,
1257 .map_fd_put_ptr = perf_event_fd_array_put_ptr,
1258 .map_release = perf_event_fd_array_release,
1259 .map_check_btf = map_check_no_btf,
1260 .map_btf_id = &array_map_btf_ids[0],
1261};
1262
1263#ifdef CONFIG_CGROUPS
1264static void *cgroup_fd_array_get_ptr(struct bpf_map *map,
1265 struct file *map_file /* not used */,
1266 int fd)
1267{
1268 return cgroup_get_from_fd(fd);
1269}
1270
1271static void cgroup_fd_array_put_ptr(void *ptr)
1272{
1273 /* cgroup_put free cgrp after a rcu grace period */
1274 cgroup_put(ptr);
1275}
1276
1277static void cgroup_fd_array_free(struct bpf_map *map)
1278{
1279 bpf_fd_array_map_clear(map);
1280 fd_array_map_free(map);
1281}
1282
1283const struct bpf_map_ops cgroup_array_map_ops = {
1284 .map_meta_equal = bpf_map_meta_equal,
1285 .map_alloc_check = fd_array_map_alloc_check,
1286 .map_alloc = array_map_alloc,
1287 .map_free = cgroup_fd_array_free,
1288 .map_get_next_key = array_map_get_next_key,
1289 .map_lookup_elem = fd_array_map_lookup_elem,
1290 .map_delete_elem = fd_array_map_delete_elem,
1291 .map_fd_get_ptr = cgroup_fd_array_get_ptr,
1292 .map_fd_put_ptr = cgroup_fd_array_put_ptr,
1293 .map_check_btf = map_check_no_btf,
1294 .map_btf_id = &array_map_btf_ids[0],
1295};
1296#endif
1297
1298static struct bpf_map *array_of_map_alloc(union bpf_attr *attr)
1299{
1300 struct bpf_map *map, *inner_map_meta;
1301
1302 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1303 if (IS_ERR(inner_map_meta))
1304 return inner_map_meta;
1305
1306 map = array_map_alloc(attr);
1307 if (IS_ERR(map)) {
1308 bpf_map_meta_free(inner_map_meta);
1309 return map;
1310 }
1311
1312 map->inner_map_meta = inner_map_meta;
1313
1314 return map;
1315}
1316
1317static void array_of_map_free(struct bpf_map *map)
1318{
1319 /* map->inner_map_meta is only accessed by syscall which
1320 * is protected by fdget/fdput.
1321 */
1322 bpf_map_meta_free(map->inner_map_meta);
1323 bpf_fd_array_map_clear(map);
1324 fd_array_map_free(map);
1325}
1326
1327static void *array_of_map_lookup_elem(struct bpf_map *map, void *key)
1328{
1329 struct bpf_map **inner_map = array_map_lookup_elem(map, key);
1330
1331 if (!inner_map)
1332 return NULL;
1333
1334 return READ_ONCE(*inner_map);
1335}
1336
1337static int array_of_map_gen_lookup(struct bpf_map *map,
1338 struct bpf_insn *insn_buf)
1339{
1340 struct bpf_array *array = container_of(map, struct bpf_array, map);
1341 u32 elem_size = array->elem_size;
1342 struct bpf_insn *insn = insn_buf;
1343 const int ret = BPF_REG_0;
1344 const int map_ptr = BPF_REG_1;
1345 const int index = BPF_REG_2;
1346
1347 *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
1348 *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
1349 if (!map->bypass_spec_v1) {
1350 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 6);
1351 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
1352 } else {
1353 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
1354 }
1355 if (is_power_of_2(elem_size))
1356 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
1357 else
1358 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
1359 *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
1360 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1361 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
1362 *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
1363 *insn++ = BPF_MOV64_IMM(ret, 0);
1364
1365 return insn - insn_buf;
1366}
1367
1368const struct bpf_map_ops array_of_maps_map_ops = {
1369 .map_alloc_check = fd_array_map_alloc_check,
1370 .map_alloc = array_of_map_alloc,
1371 .map_free = array_of_map_free,
1372 .map_get_next_key = array_map_get_next_key,
1373 .map_lookup_elem = array_of_map_lookup_elem,
1374 .map_delete_elem = fd_array_map_delete_elem,
1375 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1376 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1377 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1378 .map_gen_lookup = array_of_map_gen_lookup,
1379 .map_lookup_batch = generic_map_lookup_batch,
1380 .map_update_batch = generic_map_update_batch,
1381 .map_check_btf = map_check_no_btf,
1382 .map_btf_id = &array_map_btf_ids[0],
1383};