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