projects/linux·blob v5.11-rc1 · refs

1289 lines · 33.9 KiB

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 array_map_btf_id;
629const struct bpf_map_ops array_map_ops = {
630 .map_meta_equal = array_map_meta_equal,
631 .map_alloc_check = array_map_alloc_check,
632 .map_alloc = array_map_alloc,
633 .map_free = array_map_free,
634 .map_get_next_key = array_map_get_next_key,
635 .map_lookup_elem = array_map_lookup_elem,
636 .map_update_elem = array_map_update_elem,
637 .map_delete_elem = array_map_delete_elem,
638 .map_gen_lookup = array_map_gen_lookup,
639 .map_direct_value_addr = array_map_direct_value_addr,
640 .map_direct_value_meta = array_map_direct_value_meta,
641 .map_mmap = array_map_mmap,
642 .map_seq_show_elem = array_map_seq_show_elem,
643 .map_check_btf = array_map_check_btf,
644 .map_lookup_batch = generic_map_lookup_batch,
645 .map_update_batch = generic_map_update_batch,
646 .map_btf_name = "bpf_array",
647 .map_btf_id = &array_map_btf_id,
648 .iter_seq_info = &iter_seq_info,
649};
650
651static int percpu_array_map_btf_id;
652const struct bpf_map_ops percpu_array_map_ops = {
653 .map_meta_equal = bpf_map_meta_equal,
654 .map_alloc_check = array_map_alloc_check,
655 .map_alloc = array_map_alloc,
656 .map_free = array_map_free,
657 .map_get_next_key = array_map_get_next_key,
658 .map_lookup_elem = percpu_array_map_lookup_elem,
659 .map_update_elem = array_map_update_elem,
660 .map_delete_elem = array_map_delete_elem,
661 .map_seq_show_elem = percpu_array_map_seq_show_elem,
662 .map_check_btf = array_map_check_btf,
663 .map_btf_name = "bpf_array",
664 .map_btf_id = &percpu_array_map_btf_id,
665 .iter_seq_info = &iter_seq_info,
666};
667
668static int fd_array_map_alloc_check(union bpf_attr *attr)
669{
670 /* only file descriptors can be stored in this type of map */
671 if (attr->value_size != sizeof(u32))
672 return -EINVAL;
673 /* Program read-only/write-only not supported for special maps yet. */
674 if (attr->map_flags & (BPF_F_RDONLY_PROG | BPF_F_WRONLY_PROG))
675 return -EINVAL;
676 return array_map_alloc_check(attr);
677}
678
679static void fd_array_map_free(struct bpf_map *map)
680{
681 struct bpf_array *array = container_of(map, struct bpf_array, map);
682 int i;
683
684 /* make sure it's empty */
685 for (i = 0; i < array->map.max_entries; i++)
686 BUG_ON(array->ptrs[i] != NULL);
687
688 bpf_map_area_free(array);
689}
690
691static void *fd_array_map_lookup_elem(struct bpf_map *map, void *key)
692{
693 return ERR_PTR(-EOPNOTSUPP);
694}
695
696/* only called from syscall */
697int bpf_fd_array_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
698{
699 void **elem, *ptr;
700 int ret = 0;
701
702 if (!map->ops->map_fd_sys_lookup_elem)
703 return -ENOTSUPP;
704
705 rcu_read_lock();
706 elem = array_map_lookup_elem(map, key);
707 if (elem && (ptr = READ_ONCE(*elem)))
708 *value = map->ops->map_fd_sys_lookup_elem(ptr);
709 else
710 ret = -ENOENT;
711 rcu_read_unlock();
712
713 return ret;
714}
715
716/* only called from syscall */
717int bpf_fd_array_map_update_elem(struct bpf_map *map, struct file *map_file,
718 void *key, void *value, u64 map_flags)
719{
720 struct bpf_array *array = container_of(map, struct bpf_array, map);
721 void *new_ptr, *old_ptr;
722 u32 index = *(u32 *)key, ufd;
723
724 if (map_flags != BPF_ANY)
725 return -EINVAL;
726
727 if (index >= array->map.max_entries)
728 return -E2BIG;
729
730 ufd = *(u32 *)value;
731 new_ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
732 if (IS_ERR(new_ptr))
733 return PTR_ERR(new_ptr);
734
735 if (map->ops->map_poke_run) {
736 mutex_lock(&array->aux->poke_mutex);
737 old_ptr = xchg(array->ptrs + index, new_ptr);
738 map->ops->map_poke_run(map, index, old_ptr, new_ptr);
739 mutex_unlock(&array->aux->poke_mutex);
740 } else {
741 old_ptr = xchg(array->ptrs + index, new_ptr);
742 }
743
744 if (old_ptr)
745 map->ops->map_fd_put_ptr(old_ptr);
746 return 0;
747}
748
749static int fd_array_map_delete_elem(struct bpf_map *map, void *key)
750{
751 struct bpf_array *array = container_of(map, struct bpf_array, map);
752 void *old_ptr;
753 u32 index = *(u32 *)key;
754
755 if (index >= array->map.max_entries)
756 return -E2BIG;
757
758 if (map->ops->map_poke_run) {
759 mutex_lock(&array->aux->poke_mutex);
760 old_ptr = xchg(array->ptrs + index, NULL);
761 map->ops->map_poke_run(map, index, old_ptr, NULL);
762 mutex_unlock(&array->aux->poke_mutex);
763 } else {
764 old_ptr = xchg(array->ptrs + index, NULL);
765 }
766
767 if (old_ptr) {
768 map->ops->map_fd_put_ptr(old_ptr);
769 return 0;
770 } else {
771 return -ENOENT;
772 }
773}
774
775static void *prog_fd_array_get_ptr(struct bpf_map *map,
776 struct file *map_file, int fd)
777{
778 struct bpf_array *array = container_of(map, struct bpf_array, map);
779 struct bpf_prog *prog = bpf_prog_get(fd);
780
781 if (IS_ERR(prog))
782 return prog;
783
784 if (!bpf_prog_array_compatible(array, prog)) {
785 bpf_prog_put(prog);
786 return ERR_PTR(-EINVAL);
787 }
788
789 return prog;
790}
791
792static void prog_fd_array_put_ptr(void *ptr)
793{
794 bpf_prog_put(ptr);
795}
796
797static u32 prog_fd_array_sys_lookup_elem(void *ptr)
798{
799 return ((struct bpf_prog *)ptr)->aux->id;
800}
801
802/* decrement refcnt of all bpf_progs that are stored in this map */
803static void bpf_fd_array_map_clear(struct bpf_map *map)
804{
805 struct bpf_array *array = container_of(map, struct bpf_array, map);
806 int i;
807
808 for (i = 0; i < array->map.max_entries; i++)
809 fd_array_map_delete_elem(map, &i);
810}
811
812static void prog_array_map_seq_show_elem(struct bpf_map *map, void *key,
813 struct seq_file *m)
814{
815 void **elem, *ptr;
816 u32 prog_id;
817
818 rcu_read_lock();
819
820 elem = array_map_lookup_elem(map, key);
821 if (elem) {
822 ptr = READ_ONCE(*elem);
823 if (ptr) {
824 seq_printf(m, "%u: ", *(u32 *)key);
825 prog_id = prog_fd_array_sys_lookup_elem(ptr);
826 btf_type_seq_show(map->btf, map->btf_value_type_id,
827 &prog_id, m);
828 seq_puts(m, "\n");
829 }
830 }
831
832 rcu_read_unlock();
833}
834
835struct prog_poke_elem {
836 struct list_head list;
837 struct bpf_prog_aux *aux;
838};
839
840static int prog_array_map_poke_track(struct bpf_map *map,
841 struct bpf_prog_aux *prog_aux)
842{
843 struct prog_poke_elem *elem;
844 struct bpf_array_aux *aux;
845 int ret = 0;
846
847 aux = container_of(map, struct bpf_array, map)->aux;
848 mutex_lock(&aux->poke_mutex);
849 list_for_each_entry(elem, &aux->poke_progs, list) {
850 if (elem->aux == prog_aux)
851 goto out;
852 }
853
854 elem = kmalloc(sizeof(*elem), GFP_KERNEL);
855 if (!elem) {
856 ret = -ENOMEM;
857 goto out;
858 }
859
860 INIT_LIST_HEAD(&elem->list);
861 /* We must track the program's aux info at this point in time
862 * since the program pointer itself may not be stable yet, see
863 * also comment in prog_array_map_poke_run().
864 */
865 elem->aux = prog_aux;
866
867 list_add_tail(&elem->list, &aux->poke_progs);
868out:
869 mutex_unlock(&aux->poke_mutex);
870 return ret;
871}
872
873static void prog_array_map_poke_untrack(struct bpf_map *map,
874 struct bpf_prog_aux *prog_aux)
875{
876 struct prog_poke_elem *elem, *tmp;
877 struct bpf_array_aux *aux;
878
879 aux = container_of(map, struct bpf_array, map)->aux;
880 mutex_lock(&aux->poke_mutex);
881 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
882 if (elem->aux == prog_aux) {
883 list_del_init(&elem->list);
884 kfree(elem);
885 break;
886 }
887 }
888 mutex_unlock(&aux->poke_mutex);
889}
890
891static void prog_array_map_poke_run(struct bpf_map *map, u32 key,
892 struct bpf_prog *old,
893 struct bpf_prog *new)
894{
895 u8 *old_addr, *new_addr, *old_bypass_addr;
896 struct prog_poke_elem *elem;
897 struct bpf_array_aux *aux;
898
899 aux = container_of(map, struct bpf_array, map)->aux;
900 WARN_ON_ONCE(!mutex_is_locked(&aux->poke_mutex));
901
902 list_for_each_entry(elem, &aux->poke_progs, list) {
903 struct bpf_jit_poke_descriptor *poke;
904 int i, ret;
905
906 for (i = 0; i < elem->aux->size_poke_tab; i++) {
907 poke = &elem->aux->poke_tab[i];
908
909 /* Few things to be aware of:
910 *
911 * 1) We can only ever access aux in this context, but
912 * not aux->prog since it might not be stable yet and
913 * there could be danger of use after free otherwise.
914 * 2) Initially when we start tracking aux, the program
915 * is not JITed yet and also does not have a kallsyms
916 * entry. We skip these as poke->tailcall_target_stable
917 * is not active yet. The JIT will do the final fixup
918 * before setting it stable. The various
919 * poke->tailcall_target_stable are successively
920 * activated, so tail call updates can arrive from here
921 * while JIT is still finishing its final fixup for
922 * non-activated poke entries.
923 * 3) On program teardown, the program's kallsym entry gets
924 * removed out of RCU callback, but we can only untrack
925 * from sleepable context, therefore bpf_arch_text_poke()
926 * might not see that this is in BPF text section and
927 * bails out with -EINVAL. As these are unreachable since
928 * RCU grace period already passed, we simply skip them.
929 * 4) Also programs reaching refcount of zero while patching
930 * is in progress is okay since we're protected under
931 * poke_mutex and untrack the programs before the JIT
932 * buffer is freed. When we're still in the middle of
933 * patching and suddenly kallsyms entry of the program
934 * gets evicted, we just skip the rest which is fine due
935 * to point 3).
936 * 5) Any other error happening below from bpf_arch_text_poke()
937 * is a unexpected bug.
938 */
939 if (!READ_ONCE(poke->tailcall_target_stable))
940 continue;
941 if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
942 continue;
943 if (poke->tail_call.map != map ||
944 poke->tail_call.key != key)
945 continue;
946
947 old_bypass_addr = old ? NULL : poke->bypass_addr;
948 old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
949 new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
950
951 if (new) {
952 ret = bpf_arch_text_poke(poke->tailcall_target,
953 BPF_MOD_JUMP,
954 old_addr, new_addr);
955 BUG_ON(ret < 0 && ret != -EINVAL);
956 if (!old) {
957 ret = bpf_arch_text_poke(poke->tailcall_bypass,
958 BPF_MOD_JUMP,
959 poke->bypass_addr,
960 NULL);
961 BUG_ON(ret < 0 && ret != -EINVAL);
962 }
963 } else {
964 ret = bpf_arch_text_poke(poke->tailcall_bypass,
965 BPF_MOD_JUMP,
966 old_bypass_addr,
967 poke->bypass_addr);
968 BUG_ON(ret < 0 && ret != -EINVAL);
969 /* let other CPUs finish the execution of program
970 * so that it will not possible to expose them
971 * to invalid nop, stack unwind, nop state
972 */
973 if (!ret)
974 synchronize_rcu();
975 ret = bpf_arch_text_poke(poke->tailcall_target,
976 BPF_MOD_JUMP,
977 old_addr, NULL);
978 BUG_ON(ret < 0 && ret != -EINVAL);
979 }
980 }
981 }
982}
983
984static void prog_array_map_clear_deferred(struct work_struct *work)
985{
986 struct bpf_map *map = container_of(work, struct bpf_array_aux,
987 work)->map;
988 bpf_fd_array_map_clear(map);
989 bpf_map_put(map);
990}
991
992static void prog_array_map_clear(struct bpf_map *map)
993{
994 struct bpf_array_aux *aux = container_of(map, struct bpf_array,
995 map)->aux;
996 bpf_map_inc(map);
997 schedule_work(&aux->work);
998}
999
1000static struct bpf_map *prog_array_map_alloc(union bpf_attr *attr)
1001{
1002 struct bpf_array_aux *aux;
1003 struct bpf_map *map;
1004
1005 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT);
1006 if (!aux)
1007 return ERR_PTR(-ENOMEM);
1008
1009 INIT_WORK(&aux->work, prog_array_map_clear_deferred);
1010 INIT_LIST_HEAD(&aux->poke_progs);
1011 mutex_init(&aux->poke_mutex);
1012
1013 map = array_map_alloc(attr);
1014 if (IS_ERR(map)) {
1015 kfree(aux);
1016 return map;
1017 }
1018
1019 container_of(map, struct bpf_array, map)->aux = aux;
1020 aux->map = map;
1021
1022 return map;
1023}
1024
1025static void prog_array_map_free(struct bpf_map *map)
1026{
1027 struct prog_poke_elem *elem, *tmp;
1028 struct bpf_array_aux *aux;
1029
1030 aux = container_of(map, struct bpf_array, map)->aux;
1031 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
1032 list_del_init(&elem->list);
1033 kfree(elem);
1034 }
1035 kfree(aux);
1036 fd_array_map_free(map);
1037}
1038
1039/* prog_array->aux->{type,jited} is a runtime binding.
1040 * Doing static check alone in the verifier is not enough.
1041 * Thus, prog_array_map cannot be used as an inner_map
1042 * and map_meta_equal is not implemented.
1043 */
1044static int prog_array_map_btf_id;
1045const struct bpf_map_ops prog_array_map_ops = {
1046 .map_alloc_check = fd_array_map_alloc_check,
1047 .map_alloc = prog_array_map_alloc,
1048 .map_free = prog_array_map_free,
1049 .map_poke_track = prog_array_map_poke_track,
1050 .map_poke_untrack = prog_array_map_poke_untrack,
1051 .map_poke_run = prog_array_map_poke_run,
1052 .map_get_next_key = array_map_get_next_key,
1053 .map_lookup_elem = fd_array_map_lookup_elem,
1054 .map_delete_elem = fd_array_map_delete_elem,
1055 .map_fd_get_ptr = prog_fd_array_get_ptr,
1056 .map_fd_put_ptr = prog_fd_array_put_ptr,
1057 .map_fd_sys_lookup_elem = prog_fd_array_sys_lookup_elem,
1058 .map_release_uref = prog_array_map_clear,
1059 .map_seq_show_elem = prog_array_map_seq_show_elem,
1060 .map_btf_name = "bpf_array",
1061 .map_btf_id = &prog_array_map_btf_id,
1062};
1063
1064static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
1065 struct file *map_file)
1066{
1067 struct bpf_event_entry *ee;
1068
1069 ee = kzalloc(sizeof(*ee), GFP_ATOMIC);
1070 if (ee) {
1071 ee->event = perf_file->private_data;
1072 ee->perf_file = perf_file;
1073 ee->map_file = map_file;
1074 }
1075
1076 return ee;
1077}
1078
1079static void __bpf_event_entry_free(struct rcu_head *rcu)
1080{
1081 struct bpf_event_entry *ee;
1082
1083 ee = container_of(rcu, struct bpf_event_entry, rcu);
1084 fput(ee->perf_file);
1085 kfree(ee);
1086}
1087
1088static void bpf_event_entry_free_rcu(struct bpf_event_entry *ee)
1089{
1090 call_rcu(&ee->rcu, __bpf_event_entry_free);
1091}
1092
1093static void *perf_event_fd_array_get_ptr(struct bpf_map *map,
1094 struct file *map_file, int fd)
1095{
1096 struct bpf_event_entry *ee;
1097 struct perf_event *event;
1098 struct file *perf_file;
1099 u64 value;
1100
1101 perf_file = perf_event_get(fd);
1102 if (IS_ERR(perf_file))
1103 return perf_file;
1104
1105 ee = ERR_PTR(-EOPNOTSUPP);
1106 event = perf_file->private_data;
1107 if (perf_event_read_local(event, &value, NULL, NULL) == -EOPNOTSUPP)
1108 goto err_out;
1109
1110 ee = bpf_event_entry_gen(perf_file, map_file);
1111 if (ee)
1112 return ee;
1113 ee = ERR_PTR(-ENOMEM);
1114err_out:
1115 fput(perf_file);
1116 return ee;
1117}
1118
1119static void perf_event_fd_array_put_ptr(void *ptr)
1120{
1121 bpf_event_entry_free_rcu(ptr);
1122}
1123
1124static void perf_event_fd_array_release(struct bpf_map *map,
1125 struct file *map_file)
1126{
1127 struct bpf_array *array = container_of(map, struct bpf_array, map);
1128 struct bpf_event_entry *ee;
1129 int i;
1130
1131 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1132 return;
1133
1134 rcu_read_lock();
1135 for (i = 0; i < array->map.max_entries; i++) {
1136 ee = READ_ONCE(array->ptrs[i]);
1137 if (ee && ee->map_file == map_file)
1138 fd_array_map_delete_elem(map, &i);
1139 }
1140 rcu_read_unlock();
1141}
1142
1143static void perf_event_fd_array_map_free(struct bpf_map *map)
1144{
1145 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1146 bpf_fd_array_map_clear(map);
1147 fd_array_map_free(map);
1148}
1149
1150static int perf_event_array_map_btf_id;
1151const struct bpf_map_ops perf_event_array_map_ops = {
1152 .map_meta_equal = bpf_map_meta_equal,
1153 .map_alloc_check = fd_array_map_alloc_check,
1154 .map_alloc = array_map_alloc,
1155 .map_free = perf_event_fd_array_map_free,
1156 .map_get_next_key = array_map_get_next_key,
1157 .map_lookup_elem = fd_array_map_lookup_elem,
1158 .map_delete_elem = fd_array_map_delete_elem,
1159 .map_fd_get_ptr = perf_event_fd_array_get_ptr,
1160 .map_fd_put_ptr = perf_event_fd_array_put_ptr,
1161 .map_release = perf_event_fd_array_release,
1162 .map_check_btf = map_check_no_btf,
1163 .map_btf_name = "bpf_array",
1164 .map_btf_id = &perf_event_array_map_btf_id,
1165};
1166
1167#ifdef CONFIG_CGROUPS
1168static void *cgroup_fd_array_get_ptr(struct bpf_map *map,
1169 struct file *map_file /* not used */,
1170 int fd)
1171{
1172 return cgroup_get_from_fd(fd);
1173}
1174
1175static void cgroup_fd_array_put_ptr(void *ptr)
1176{
1177 /* cgroup_put free cgrp after a rcu grace period */
1178 cgroup_put(ptr);
1179}
1180
1181static void cgroup_fd_array_free(struct bpf_map *map)
1182{
1183 bpf_fd_array_map_clear(map);
1184 fd_array_map_free(map);
1185}
1186
1187static int cgroup_array_map_btf_id;
1188const struct bpf_map_ops cgroup_array_map_ops = {
1189 .map_meta_equal = bpf_map_meta_equal,
1190 .map_alloc_check = fd_array_map_alloc_check,
1191 .map_alloc = array_map_alloc,
1192 .map_free = cgroup_fd_array_free,
1193 .map_get_next_key = array_map_get_next_key,
1194 .map_lookup_elem = fd_array_map_lookup_elem,
1195 .map_delete_elem = fd_array_map_delete_elem,
1196 .map_fd_get_ptr = cgroup_fd_array_get_ptr,
1197 .map_fd_put_ptr = cgroup_fd_array_put_ptr,
1198 .map_check_btf = map_check_no_btf,
1199 .map_btf_name = "bpf_array",
1200 .map_btf_id = &cgroup_array_map_btf_id,
1201};
1202#endif
1203
1204static struct bpf_map *array_of_map_alloc(union bpf_attr *attr)
1205{
1206 struct bpf_map *map, *inner_map_meta;
1207
1208 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1209 if (IS_ERR(inner_map_meta))
1210 return inner_map_meta;
1211
1212 map = array_map_alloc(attr);
1213 if (IS_ERR(map)) {
1214 bpf_map_meta_free(inner_map_meta);
1215 return map;
1216 }
1217
1218 map->inner_map_meta = inner_map_meta;
1219
1220 return map;
1221}
1222
1223static void array_of_map_free(struct bpf_map *map)
1224{
1225 /* map->inner_map_meta is only accessed by syscall which
1226 * is protected by fdget/fdput.
1227 */
1228 bpf_map_meta_free(map->inner_map_meta);
1229 bpf_fd_array_map_clear(map);
1230 fd_array_map_free(map);
1231}
1232
1233static void *array_of_map_lookup_elem(struct bpf_map *map, void *key)
1234{
1235 struct bpf_map **inner_map = array_map_lookup_elem(map, key);
1236
1237 if (!inner_map)
1238 return NULL;
1239
1240 return READ_ONCE(*inner_map);
1241}
1242
1243static int array_of_map_gen_lookup(struct bpf_map *map,
1244 struct bpf_insn *insn_buf)
1245{
1246 struct bpf_array *array = container_of(map, struct bpf_array, map);
1247 u32 elem_size = round_up(map->value_size, 8);
1248 struct bpf_insn *insn = insn_buf;
1249 const int ret = BPF_REG_0;
1250 const int map_ptr = BPF_REG_1;
1251 const int index = BPF_REG_2;
1252
1253 *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
1254 *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
1255 if (!map->bypass_spec_v1) {
1256 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 6);
1257 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
1258 } else {
1259 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
1260 }
1261 if (is_power_of_2(elem_size))
1262 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
1263 else
1264 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
1265 *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
1266 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1267 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
1268 *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
1269 *insn++ = BPF_MOV64_IMM(ret, 0);
1270
1271 return insn - insn_buf;
1272}
1273
1274static int array_of_maps_map_btf_id;
1275const struct bpf_map_ops array_of_maps_map_ops = {
1276 .map_alloc_check = fd_array_map_alloc_check,
1277 .map_alloc = array_of_map_alloc,
1278 .map_free = array_of_map_free,
1279 .map_get_next_key = array_map_get_next_key,
1280 .map_lookup_elem = array_of_map_lookup_elem,
1281 .map_delete_elem = fd_array_map_delete_elem,
1282 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1283 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1284 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1285 .map_gen_lookup = array_of_map_gen_lookup,
1286 .map_check_btf = map_check_no_btf,
1287 .map_btf_name = "bpf_array",
1288 .map_btf_id = &array_of_maps_map_btf_id,
1289};