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