1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2019 Facebook */
3
4 #include <linux/bpf.h>
5 #include <linux/bpf_verifier.h>
6 #include <linux/btf.h>
7 #include <linux/filter.h>
8 #include <linux/slab.h>
9 #include <linux/numa.h>
10 #include <linux/seq_file.h>
11 #include <linux/refcount.h>
12 #include <linux/mutex.h>
13 #include <linux/btf_ids.h>
14 #include <linux/rcupdate_wait.h>
15 #include <linux/poll.h>
16
17 struct bpf_struct_ops_value {
18 struct bpf_struct_ops_common_value common;
19 char data[] ____cacheline_aligned_in_smp;
20 };
21
22 #define MAX_TRAMP_IMAGE_PAGES 8
23
24 struct bpf_struct_ops_map {
25 struct bpf_map map;
26 const struct bpf_struct_ops_desc *st_ops_desc;
27 /* protect map_update */
28 struct mutex lock;
29 /* link has all the bpf_links that is populated
30 * to the func ptr of the kernel's struct
31 * (in kvalue.data).
32 */
33 struct bpf_link **links;
34 /* ksyms for bpf trampolines */
35 struct bpf_ksym **ksyms;
36 u32 funcs_cnt;
37 u32 image_pages_cnt;
38 /* image_pages is an array of pages that has all the trampolines
39 * that stores the func args before calling the bpf_prog.
40 */
41 void *image_pages[MAX_TRAMP_IMAGE_PAGES];
42 /* The owner moduler's btf. */
43 struct btf *btf;
44 /* uvalue->data stores the kernel struct
45 * (e.g. tcp_congestion_ops) that is more useful
46 * to userspace than the kvalue. For example,
47 * the bpf_prog's id is stored instead of the kernel
48 * address of a func ptr.
49 */
50 struct bpf_struct_ops_value *uvalue;
51 /* kvalue.data stores the actual kernel's struct
52 * (e.g. tcp_congestion_ops) that will be
53 * registered to the kernel subsystem.
54 */
55 struct bpf_struct_ops_value kvalue;
56 };
57
58 struct bpf_struct_ops_link {
59 struct bpf_link link;
60 struct bpf_map __rcu *map;
61 wait_queue_head_t wait_hup;
62 };
63
64 static DEFINE_MUTEX(update_mutex);
65
66 #define VALUE_PREFIX "bpf_struct_ops_"
67 #define VALUE_PREFIX_LEN (sizeof(VALUE_PREFIX) - 1)
68
69 const struct bpf_verifier_ops bpf_struct_ops_verifier_ops = {
70 };
71
72 const struct bpf_prog_ops bpf_struct_ops_prog_ops = {
73 #ifdef CONFIG_NET
74 .test_run = bpf_struct_ops_test_run,
75 #endif
76 };
77
78 BTF_ID_LIST(st_ops_ids)
79 BTF_ID(struct, module)
80 BTF_ID(struct, bpf_struct_ops_common_value)
81
82 enum {
83 IDX_MODULE_ID,
84 IDX_ST_OPS_COMMON_VALUE_ID,
85 };
86
87 extern struct btf *btf_vmlinux;
88
is_valid_value_type(struct btf * btf,s32 value_id,const struct btf_type * type,const char * value_name)89 static bool is_valid_value_type(struct btf *btf, s32 value_id,
90 const struct btf_type *type,
91 const char *value_name)
92 {
93 const struct btf_type *common_value_type;
94 const struct btf_member *member;
95 const struct btf_type *vt, *mt;
96
97 vt = btf_type_by_id(btf, value_id);
98 if (btf_vlen(vt) != 2) {
99 pr_warn("The number of %s's members should be 2, but we get %d\n",
100 value_name, btf_vlen(vt));
101 return false;
102 }
103 member = btf_type_member(vt);
104 mt = btf_type_by_id(btf, member->type);
105 common_value_type = btf_type_by_id(btf_vmlinux,
106 st_ops_ids[IDX_ST_OPS_COMMON_VALUE_ID]);
107 if (mt != common_value_type) {
108 pr_warn("The first member of %s should be bpf_struct_ops_common_value\n",
109 value_name);
110 return false;
111 }
112 member++;
113 mt = btf_type_by_id(btf, member->type);
114 if (mt != type) {
115 pr_warn("The second member of %s should be %s\n",
116 value_name, btf_name_by_offset(btf, type->name_off));
117 return false;
118 }
119
120 return true;
121 }
122
bpf_struct_ops_image_alloc(void)123 static void *bpf_struct_ops_image_alloc(void)
124 {
125 void *image;
126 int err;
127
128 err = bpf_jit_charge_modmem(PAGE_SIZE);
129 if (err)
130 return ERR_PTR(err);
131 image = arch_alloc_bpf_trampoline(PAGE_SIZE);
132 if (!image) {
133 bpf_jit_uncharge_modmem(PAGE_SIZE);
134 return ERR_PTR(-ENOMEM);
135 }
136
137 return image;
138 }
139
bpf_struct_ops_image_free(void * image)140 void bpf_struct_ops_image_free(void *image)
141 {
142 if (image) {
143 arch_free_bpf_trampoline(image, PAGE_SIZE);
144 bpf_jit_uncharge_modmem(PAGE_SIZE);
145 }
146 }
147
148 #define MAYBE_NULL_SUFFIX "__nullable"
149 #define REFCOUNTED_SUFFIX "__ref"
150
151 /* Prepare argument info for every nullable argument of a member of a
152 * struct_ops type.
153 *
154 * Initialize a struct bpf_struct_ops_arg_info according to type info of
155 * the arguments of a stub function. (Check kCFI for more information about
156 * stub functions.)
157 *
158 * Each member in the struct_ops type has a struct bpf_struct_ops_arg_info
159 * to provide an array of struct bpf_ctx_arg_aux, which in turn provides
160 * the information that used by the verifier to check the arguments of the
161 * BPF struct_ops program assigned to the member. Here, we only care about
162 * the arguments that are marked as __nullable.
163 *
164 * The array of struct bpf_ctx_arg_aux is eventually assigned to
165 * prog->aux->ctx_arg_info of BPF struct_ops programs and passed to the
166 * verifier. (See check_struct_ops_btf_id())
167 *
168 * arg_info->info will be the list of struct bpf_ctx_arg_aux if success. If
169 * fails, it will be kept untouched.
170 */
prepare_arg_info(struct btf * btf,const char * st_ops_name,const char * member_name,const struct btf_type * func_proto,void * stub_func_addr,struct bpf_struct_ops_arg_info * arg_info)171 static int prepare_arg_info(struct btf *btf,
172 const char *st_ops_name,
173 const char *member_name,
174 const struct btf_type *func_proto, void *stub_func_addr,
175 struct bpf_struct_ops_arg_info *arg_info)
176 {
177 const struct btf_type *stub_func_proto, *pointed_type;
178 bool is_nullable = false, is_refcounted = false;
179 const struct btf_param *stub_args, *args;
180 struct bpf_ctx_arg_aux *info, *info_buf;
181 u32 nargs, arg_no, info_cnt = 0;
182 char ksym[KSYM_SYMBOL_LEN];
183 const char *stub_fname;
184 const char *suffix;
185 s32 stub_func_id;
186 u32 arg_btf_id;
187 int offset;
188
189 stub_fname = kallsyms_lookup((unsigned long)stub_func_addr, NULL, NULL, NULL, ksym);
190 if (!stub_fname) {
191 pr_warn("Cannot find the stub function name for the %s in struct %s\n",
192 member_name, st_ops_name);
193 return -ENOENT;
194 }
195
196 stub_func_id = btf_find_by_name_kind(btf, stub_fname, BTF_KIND_FUNC);
197 if (stub_func_id < 0) {
198 pr_warn("Cannot find the stub function %s in btf\n", stub_fname);
199 return -ENOENT;
200 }
201
202 stub_func_proto = btf_type_by_id(btf, stub_func_id);
203 stub_func_proto = btf_type_by_id(btf, stub_func_proto->type);
204
205 /* Check if the number of arguments of the stub function is the same
206 * as the number of arguments of the function pointer.
207 */
208 nargs = btf_type_vlen(func_proto);
209 if (nargs != btf_type_vlen(stub_func_proto)) {
210 pr_warn("the number of arguments of the stub function %s does not match the number of arguments of the member %s of struct %s\n",
211 stub_fname, member_name, st_ops_name);
212 return -EINVAL;
213 }
214
215 if (!nargs)
216 return 0;
217
218 args = btf_params(func_proto);
219 stub_args = btf_params(stub_func_proto);
220
221 info_buf = kzalloc_objs(*info_buf, nargs, GFP_KERNEL);
222 if (!info_buf)
223 return -ENOMEM;
224
225 /* Prepare info for every nullable argument */
226 info = info_buf;
227 for (arg_no = 0; arg_no < nargs; arg_no++) {
228 /* Skip arguments that is not suffixed with
229 * "__nullable or __ref".
230 */
231 is_nullable = btf_param_match_suffix(btf, &stub_args[arg_no],
232 MAYBE_NULL_SUFFIX);
233 is_refcounted = btf_param_match_suffix(btf, &stub_args[arg_no],
234 REFCOUNTED_SUFFIX);
235
236 if (is_nullable)
237 suffix = MAYBE_NULL_SUFFIX;
238 else if (is_refcounted)
239 suffix = REFCOUNTED_SUFFIX;
240 else
241 continue;
242
243 /* Should be a pointer to struct */
244 pointed_type = btf_type_resolve_ptr(btf,
245 args[arg_no].type,
246 &arg_btf_id);
247 if (!pointed_type ||
248 !btf_type_is_struct(pointed_type)) {
249 pr_warn("stub function %s has %s tagging to an unsupported type\n",
250 stub_fname, suffix);
251 goto err_out;
252 }
253
254 offset = btf_ctx_arg_offset(btf, func_proto, arg_no);
255 if (offset < 0) {
256 pr_warn("stub function %s has an invalid trampoline ctx offset for arg#%u\n",
257 stub_fname, arg_no);
258 goto err_out;
259 }
260
261 if (args[arg_no].type != stub_args[arg_no].type) {
262 pr_warn("arg#%u type in stub function %s does not match with its original func_proto\n",
263 arg_no, stub_fname);
264 goto err_out;
265 }
266
267 /* Fill the information of the new argument */
268 info->btf_id = arg_btf_id;
269 info->btf = btf;
270 info->offset = offset;
271 if (is_nullable) {
272 info->reg_type = PTR_TRUSTED | PTR_TO_BTF_ID | PTR_MAYBE_NULL;
273 } else if (is_refcounted) {
274 info->reg_type = PTR_TRUSTED | PTR_TO_BTF_ID;
275 info->refcounted = true;
276 }
277
278 info++;
279 info_cnt++;
280 }
281
282 if (info_cnt) {
283 arg_info->info = info_buf;
284 arg_info->cnt = info_cnt;
285 } else {
286 kfree(info_buf);
287 }
288
289 return 0;
290
291 err_out:
292 kfree(info_buf);
293
294 return -EINVAL;
295 }
296
297 /* Clean up the arg_info in a struct bpf_struct_ops_desc. */
bpf_struct_ops_desc_release(struct bpf_struct_ops_desc * st_ops_desc)298 void bpf_struct_ops_desc_release(struct bpf_struct_ops_desc *st_ops_desc)
299 {
300 struct bpf_struct_ops_arg_info *arg_info;
301 int i;
302
303 arg_info = st_ops_desc->arg_info;
304 for (i = 0; i < btf_type_vlen(st_ops_desc->type); i++)
305 kfree(arg_info[i].info);
306
307 kfree(arg_info);
308 }
309
is_module_member(const struct btf * btf,u32 id)310 static bool is_module_member(const struct btf *btf, u32 id)
311 {
312 const struct btf_type *t;
313
314 t = btf_type_resolve_ptr(btf, id, NULL);
315 if (!t)
316 return false;
317
318 if (!__btf_type_is_struct(t) && !btf_type_is_fwd(t))
319 return false;
320
321 return !strcmp(btf_name_by_offset(btf, t->name_off), "module");
322 }
323
bpf_struct_ops_supported(const struct bpf_struct_ops * st_ops,u32 moff)324 int bpf_struct_ops_supported(const struct bpf_struct_ops *st_ops, u32 moff)
325 {
326 void *func_ptr = *(void **)(st_ops->cfi_stubs + moff);
327
328 return func_ptr ? 0 : -ENOTSUPP;
329 }
330
bpf_struct_ops_desc_init(struct bpf_struct_ops_desc * st_ops_desc,struct btf * btf,struct bpf_verifier_log * log)331 int bpf_struct_ops_desc_init(struct bpf_struct_ops_desc *st_ops_desc,
332 struct btf *btf,
333 struct bpf_verifier_log *log)
334 {
335 struct bpf_struct_ops *st_ops = st_ops_desc->st_ops;
336 struct bpf_struct_ops_arg_info *arg_info;
337 const struct btf_member *member;
338 const struct btf_type *t;
339 s32 type_id, value_id;
340 char value_name[128];
341 const char *mname;
342 int i, err;
343
344 if (strlen(st_ops->name) + VALUE_PREFIX_LEN >=
345 sizeof(value_name)) {
346 pr_warn("struct_ops name %s is too long\n",
347 st_ops->name);
348 return -EINVAL;
349 }
350 sprintf(value_name, "%s%s", VALUE_PREFIX, st_ops->name);
351
352 if (!st_ops->cfi_stubs) {
353 pr_warn("struct_ops for %s has no cfi_stubs\n", st_ops->name);
354 return -EINVAL;
355 }
356
357 type_id = btf_find_by_name_kind(btf, st_ops->name,
358 BTF_KIND_STRUCT);
359 if (type_id < 0) {
360 pr_warn("Cannot find struct %s in %s\n",
361 st_ops->name, btf_get_name(btf));
362 return -EINVAL;
363 }
364 t = btf_type_by_id(btf, type_id);
365 if (btf_type_vlen(t) > BPF_STRUCT_OPS_MAX_NR_MEMBERS) {
366 pr_warn("Cannot support #%u members in struct %s\n",
367 btf_type_vlen(t), st_ops->name);
368 return -EINVAL;
369 }
370
371 value_id = btf_find_by_name_kind(btf, value_name,
372 BTF_KIND_STRUCT);
373 if (value_id < 0) {
374 pr_warn("Cannot find struct %s in %s\n",
375 value_name, btf_get_name(btf));
376 return -EINVAL;
377 }
378 if (!is_valid_value_type(btf, value_id, t, value_name))
379 return -EINVAL;
380
381 arg_info = kzalloc_objs(*arg_info, btf_type_vlen(t), GFP_KERNEL);
382 if (!arg_info)
383 return -ENOMEM;
384
385 st_ops_desc->arg_info = arg_info;
386 st_ops_desc->type = t;
387 st_ops_desc->type_id = type_id;
388 st_ops_desc->value_id = value_id;
389 st_ops_desc->value_type = btf_type_by_id(btf, value_id);
390
391 for_each_member(i, t, member) {
392 const struct btf_type *func_proto, *ret_type;
393 void **stub_func_addr;
394 u32 moff;
395
396 moff = __btf_member_bit_offset(t, member) / 8;
397 mname = btf_name_by_offset(btf, member->name_off);
398 if (!*mname) {
399 pr_warn("anon member in struct %s is not supported\n",
400 st_ops->name);
401 err = -EOPNOTSUPP;
402 goto errout;
403 }
404
405 if (__btf_member_bitfield_size(t, member)) {
406 pr_warn("bit field member %s in struct %s is not supported\n",
407 mname, st_ops->name);
408 err = -EOPNOTSUPP;
409 goto errout;
410 }
411
412 if (!st_ops_ids[IDX_MODULE_ID] && is_module_member(btf, member->type)) {
413 pr_warn("'struct module' btf id not found. Is CONFIG_MODULES enabled? bpf_struct_ops '%s' needs module support.\n",
414 st_ops->name);
415 err = -EOPNOTSUPP;
416 goto errout;
417 }
418
419 func_proto = btf_type_resolve_func_ptr(btf,
420 member->type,
421 NULL);
422
423 /* The member is not a function pointer or
424 * the function pointer is not supported.
425 */
426 if (!func_proto || bpf_struct_ops_supported(st_ops, moff))
427 continue;
428
429 if (func_proto->type) {
430 ret_type = btf_type_resolve_ptr(btf, func_proto->type, NULL);
431 if (ret_type && !__btf_type_is_struct(ret_type)) {
432 pr_warn("func ptr %s in struct %s returns non-struct pointer, which is not supported\n",
433 mname, st_ops->name);
434 err = -EOPNOTSUPP;
435 goto errout;
436 }
437 }
438
439 if (btf_distill_func_proto(log, btf,
440 func_proto, mname,
441 &st_ops->func_models[i])) {
442 pr_warn("Error in parsing func ptr %s in struct %s\n",
443 mname, st_ops->name);
444 err = -EINVAL;
445 goto errout;
446 }
447
448 stub_func_addr = *(void **)(st_ops->cfi_stubs + moff);
449 err = prepare_arg_info(btf, st_ops->name, mname,
450 func_proto, stub_func_addr,
451 arg_info + i);
452 if (err)
453 goto errout;
454 }
455
456 if (st_ops->init(btf)) {
457 pr_warn("Error in init bpf_struct_ops %s\n",
458 st_ops->name);
459 err = -EINVAL;
460 goto errout;
461 }
462
463 return 0;
464
465 errout:
466 bpf_struct_ops_desc_release(st_ops_desc);
467
468 return err;
469 }
470
bpf_struct_ops_map_get_next_key(struct bpf_map * map,void * key,void * next_key)471 static int bpf_struct_ops_map_get_next_key(struct bpf_map *map, void *key,
472 void *next_key)
473 {
474 if (key && *(u32 *)key == 0)
475 return -ENOENT;
476
477 *(u32 *)next_key = 0;
478 return 0;
479 }
480
bpf_struct_ops_map_sys_lookup_elem(struct bpf_map * map,void * key,void * value)481 int bpf_struct_ops_map_sys_lookup_elem(struct bpf_map *map, void *key,
482 void *value)
483 {
484 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
485 struct bpf_struct_ops_value *uvalue, *kvalue;
486 enum bpf_struct_ops_state state;
487 s64 refcnt;
488
489 if (unlikely(*(u32 *)key != 0))
490 return -ENOENT;
491
492 kvalue = &st_map->kvalue;
493 /* Pair with smp_store_release() during map_update */
494 state = smp_load_acquire(&kvalue->common.state);
495 if (state == BPF_STRUCT_OPS_STATE_INIT) {
496 memset(value, 0, map->value_size);
497 return 0;
498 }
499
500 /* No lock is needed. state and refcnt do not need
501 * to be updated together under atomic context.
502 */
503 uvalue = value;
504 memcpy(uvalue, st_map->uvalue, map->value_size);
505 uvalue->common.state = state;
506
507 /* This value offers the user space a general estimate of how
508 * many sockets are still utilizing this struct_ops for TCP
509 * congestion control. The number might not be exact, but it
510 * should sufficiently meet our present goals.
511 */
512 refcnt = atomic64_read(&map->refcnt) - atomic64_read(&map->usercnt);
513 refcount_set(&uvalue->common.refcnt, max_t(s64, refcnt, 0));
514
515 return 0;
516 }
517
bpf_struct_ops_map_lookup_elem(struct bpf_map * map,void * key)518 static void *bpf_struct_ops_map_lookup_elem(struct bpf_map *map, void *key)
519 {
520 return ERR_PTR(-EINVAL);
521 }
522
bpf_struct_ops_map_put_progs(struct bpf_struct_ops_map * st_map)523 static void bpf_struct_ops_map_put_progs(struct bpf_struct_ops_map *st_map)
524 {
525 u32 i;
526
527 for (i = 0; i < st_map->funcs_cnt; i++) {
528 if (!st_map->links[i])
529 break;
530 bpf_link_put(st_map->links[i]);
531 st_map->links[i] = NULL;
532 }
533 }
534
bpf_struct_ops_map_dissoc_progs(struct bpf_struct_ops_map * st_map)535 static void bpf_struct_ops_map_dissoc_progs(struct bpf_struct_ops_map *st_map)
536 {
537 u32 i;
538
539 for (i = 0; i < st_map->funcs_cnt; i++) {
540 if (!st_map->links[i])
541 break;
542 bpf_prog_disassoc_struct_ops(st_map->links[i]->prog);
543 }
544 }
545
bpf_struct_ops_map_free_image(struct bpf_struct_ops_map * st_map)546 static void bpf_struct_ops_map_free_image(struct bpf_struct_ops_map *st_map)
547 {
548 int i;
549
550 for (i = 0; i < st_map->image_pages_cnt; i++)
551 bpf_struct_ops_image_free(st_map->image_pages[i]);
552 st_map->image_pages_cnt = 0;
553 }
554
check_zero_holes(const struct btf * btf,const struct btf_type * t,void * data)555 static int check_zero_holes(const struct btf *btf, const struct btf_type *t, void *data)
556 {
557 const struct btf_member *member;
558 u32 i, moff, msize, prev_mend = 0;
559 const struct btf_type *mtype;
560
561 for_each_member(i, t, member) {
562 moff = __btf_member_bit_offset(t, member) / 8;
563 if (moff > prev_mend &&
564 memchr_inv(data + prev_mend, 0, moff - prev_mend))
565 return -EINVAL;
566
567 mtype = btf_type_by_id(btf, member->type);
568 mtype = btf_resolve_size(btf, mtype, &msize);
569 if (IS_ERR(mtype))
570 return PTR_ERR(mtype);
571 prev_mend = moff + msize;
572 }
573
574 if (t->size > prev_mend &&
575 memchr_inv(data + prev_mend, 0, t->size - prev_mend))
576 return -EINVAL;
577
578 return 0;
579 }
580
bpf_struct_ops_link_release(struct bpf_link * link)581 static void bpf_struct_ops_link_release(struct bpf_link *link)
582 {
583 }
584
bpf_struct_ops_link_dealloc(struct bpf_link * link)585 static void bpf_struct_ops_link_dealloc(struct bpf_link *link)
586 {
587 struct bpf_tramp_link *tlink = container_of(link, struct bpf_tramp_link, link);
588
589 kfree(tlink);
590 }
591
592 const struct bpf_link_ops bpf_struct_ops_link_lops = {
593 .release = bpf_struct_ops_link_release,
594 .dealloc = bpf_struct_ops_link_dealloc,
595 };
596
bpf_struct_ops_prepare_trampoline(struct bpf_tramp_links * tlinks,struct bpf_tramp_link * link,const struct btf_func_model * model,void * stub_func,void ** _image,u32 * _image_off,bool allow_alloc)597 int bpf_struct_ops_prepare_trampoline(struct bpf_tramp_links *tlinks,
598 struct bpf_tramp_link *link,
599 const struct btf_func_model *model,
600 void *stub_func,
601 void **_image, u32 *_image_off,
602 bool allow_alloc)
603 {
604 u32 image_off = *_image_off, flags = BPF_TRAMP_F_INDIRECT;
605 void *image = *_image;
606 int size;
607
608 tlinks[BPF_TRAMP_FENTRY].links[0] = link;
609 tlinks[BPF_TRAMP_FENTRY].nr_links = 1;
610
611 if (model->ret_size > 0)
612 flags |= BPF_TRAMP_F_RET_FENTRY_RET;
613
614 size = arch_bpf_trampoline_size(model, flags, tlinks, stub_func);
615 if (size <= 0)
616 return size ? : -EFAULT;
617
618 /* Allocate image buffer if necessary */
619 if (!image || size > PAGE_SIZE - image_off) {
620 if (!allow_alloc)
621 return -E2BIG;
622
623 image = bpf_struct_ops_image_alloc();
624 if (IS_ERR(image))
625 return PTR_ERR(image);
626 image_off = 0;
627 }
628
629 size = arch_prepare_bpf_trampoline(NULL, image + image_off,
630 image + image_off + size,
631 model, flags, tlinks, stub_func);
632 if (size <= 0) {
633 if (image != *_image)
634 bpf_struct_ops_image_free(image);
635 return size ? : -EFAULT;
636 }
637
638 *_image = image;
639 *_image_off = image_off + size;
640 return 0;
641 }
642
bpf_struct_ops_ksym_init(const char * tname,const char * mname,void * image,unsigned int size,struct bpf_ksym * ksym)643 static void bpf_struct_ops_ksym_init(const char *tname, const char *mname,
644 void *image, unsigned int size,
645 struct bpf_ksym *ksym)
646 {
647 snprintf(ksym->name, KSYM_NAME_LEN, "bpf__%s_%s", tname, mname);
648 INIT_LIST_HEAD_RCU(&ksym->lnode);
649 bpf_image_ksym_init(image, size, ksym);
650 }
651
bpf_struct_ops_map_add_ksyms(struct bpf_struct_ops_map * st_map)652 static void bpf_struct_ops_map_add_ksyms(struct bpf_struct_ops_map *st_map)
653 {
654 u32 i;
655
656 for (i = 0; i < st_map->funcs_cnt; i++) {
657 if (!st_map->ksyms[i])
658 break;
659 bpf_image_ksym_add(st_map->ksyms[i]);
660 }
661 }
662
bpf_struct_ops_map_del_ksyms(struct bpf_struct_ops_map * st_map)663 static void bpf_struct_ops_map_del_ksyms(struct bpf_struct_ops_map *st_map)
664 {
665 u32 i;
666
667 for (i = 0; i < st_map->funcs_cnt; i++) {
668 if (!st_map->ksyms[i])
669 break;
670 bpf_image_ksym_del(st_map->ksyms[i]);
671 }
672 }
673
bpf_struct_ops_map_free_ksyms(struct bpf_struct_ops_map * st_map)674 static void bpf_struct_ops_map_free_ksyms(struct bpf_struct_ops_map *st_map)
675 {
676 u32 i;
677
678 for (i = 0; i < st_map->funcs_cnt; i++) {
679 if (!st_map->ksyms[i])
680 break;
681 kfree(st_map->ksyms[i]);
682 st_map->ksyms[i] = NULL;
683 }
684 }
685
bpf_struct_ops_map_update_elem(struct bpf_map * map,void * key,void * value,u64 flags)686 static long bpf_struct_ops_map_update_elem(struct bpf_map *map, void *key,
687 void *value, u64 flags)
688 {
689 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
690 const struct bpf_struct_ops_desc *st_ops_desc = st_map->st_ops_desc;
691 const struct bpf_struct_ops *st_ops = st_ops_desc->st_ops;
692 struct bpf_struct_ops_value *uvalue, *kvalue;
693 const struct btf_type *module_type;
694 const struct btf_member *member;
695 const struct btf_type *t = st_ops_desc->type;
696 struct bpf_tramp_links *tlinks;
697 void *udata, *kdata;
698 int prog_fd, err;
699 u32 i, trampoline_start, image_off = 0;
700 void *cur_image = NULL, *image = NULL;
701 struct bpf_link **plink;
702 struct bpf_ksym **pksym;
703 const char *tname, *mname;
704
705 if (flags)
706 return -EINVAL;
707
708 if (*(u32 *)key != 0)
709 return -E2BIG;
710
711 err = check_zero_holes(st_map->btf, st_ops_desc->value_type, value);
712 if (err)
713 return err;
714
715 uvalue = value;
716 err = check_zero_holes(st_map->btf, t, uvalue->data);
717 if (err)
718 return err;
719
720 if (uvalue->common.state || refcount_read(&uvalue->common.refcnt))
721 return -EINVAL;
722
723 tlinks = kzalloc_objs(*tlinks, BPF_TRAMP_MAX, GFP_KERNEL);
724 if (!tlinks)
725 return -ENOMEM;
726
727 uvalue = (struct bpf_struct_ops_value *)st_map->uvalue;
728 kvalue = (struct bpf_struct_ops_value *)&st_map->kvalue;
729
730 mutex_lock(&st_map->lock);
731
732 if (kvalue->common.state != BPF_STRUCT_OPS_STATE_INIT) {
733 err = -EBUSY;
734 goto unlock;
735 }
736
737 memcpy(uvalue, value, map->value_size);
738
739 udata = &uvalue->data;
740 kdata = &kvalue->data;
741
742 plink = st_map->links;
743 pksym = st_map->ksyms;
744 tname = btf_name_by_offset(st_map->btf, t->name_off);
745 module_type = btf_type_by_id(btf_vmlinux, st_ops_ids[IDX_MODULE_ID]);
746 for_each_member(i, t, member) {
747 const struct btf_type *mtype, *ptype;
748 struct bpf_prog *prog;
749 struct bpf_tramp_link *link;
750 struct bpf_ksym *ksym;
751 u32 moff;
752
753 moff = __btf_member_bit_offset(t, member) / 8;
754 mname = btf_name_by_offset(st_map->btf, member->name_off);
755 ptype = btf_type_resolve_ptr(st_map->btf, member->type, NULL);
756 if (ptype == module_type) {
757 if (*(void **)(udata + moff))
758 goto reset_unlock;
759 *(void **)(kdata + moff) = BPF_MODULE_OWNER;
760 continue;
761 }
762
763 err = st_ops->init_member(t, member, kdata, udata);
764 if (err < 0)
765 goto reset_unlock;
766
767 /* The ->init_member() has handled this member */
768 if (err > 0)
769 continue;
770
771 /* If st_ops->init_member does not handle it,
772 * we will only handle func ptrs and zero-ed members
773 * here. Reject everything else.
774 */
775
776 /* All non func ptr member must be 0 */
777 if (!ptype || !btf_type_is_func_proto(ptype)) {
778 u32 msize;
779
780 mtype = btf_type_by_id(st_map->btf, member->type);
781 mtype = btf_resolve_size(st_map->btf, mtype, &msize);
782 if (IS_ERR(mtype)) {
783 err = PTR_ERR(mtype);
784 goto reset_unlock;
785 }
786
787 if (memchr_inv(udata + moff, 0, msize)) {
788 err = -EINVAL;
789 goto reset_unlock;
790 }
791
792 continue;
793 }
794
795 prog_fd = (int)(*(unsigned long *)(udata + moff));
796 /* Similar check as the attr->attach_prog_fd */
797 if (!prog_fd)
798 continue;
799
800 prog = bpf_prog_get(prog_fd);
801 if (IS_ERR(prog)) {
802 err = PTR_ERR(prog);
803 goto reset_unlock;
804 }
805
806 if (prog->type != BPF_PROG_TYPE_STRUCT_OPS ||
807 prog->aux->attach_btf_id != st_ops_desc->type_id ||
808 prog->expected_attach_type != i) {
809 bpf_prog_put(prog);
810 err = -EINVAL;
811 goto reset_unlock;
812 }
813
814 /* Poison pointer on error instead of return for backward compatibility */
815 bpf_prog_assoc_struct_ops(prog, &st_map->map);
816
817 link = kzalloc_obj(*link, GFP_USER);
818 if (!link) {
819 bpf_prog_put(prog);
820 err = -ENOMEM;
821 goto reset_unlock;
822 }
823 bpf_link_init(&link->link, BPF_LINK_TYPE_STRUCT_OPS,
824 &bpf_struct_ops_link_lops, prog, prog->expected_attach_type);
825 *plink++ = &link->link;
826
827 ksym = kzalloc_obj(*ksym, GFP_USER);
828 if (!ksym) {
829 err = -ENOMEM;
830 goto reset_unlock;
831 }
832 *pksym++ = ksym;
833
834 trampoline_start = image_off;
835 err = bpf_struct_ops_prepare_trampoline(tlinks, link,
836 &st_ops->func_models[i],
837 *(void **)(st_ops->cfi_stubs + moff),
838 &image, &image_off,
839 st_map->image_pages_cnt < MAX_TRAMP_IMAGE_PAGES);
840 if (err)
841 goto reset_unlock;
842
843 if (cur_image != image) {
844 st_map->image_pages[st_map->image_pages_cnt++] = image;
845 cur_image = image;
846 trampoline_start = 0;
847 }
848
849 *(void **)(kdata + moff) = image + trampoline_start + cfi_get_offset();
850
851 /* put prog_id to udata */
852 *(unsigned long *)(udata + moff) = prog->aux->id;
853
854 /* init ksym for this trampoline */
855 bpf_struct_ops_ksym_init(tname, mname,
856 image + trampoline_start,
857 image_off - trampoline_start,
858 ksym);
859 }
860
861 if (st_ops->validate) {
862 err = st_ops->validate(kdata);
863 if (err)
864 goto reset_unlock;
865 }
866 for (i = 0; i < st_map->image_pages_cnt; i++) {
867 err = arch_protect_bpf_trampoline(st_map->image_pages[i],
868 PAGE_SIZE);
869 if (err)
870 goto reset_unlock;
871 }
872
873 if (st_map->map.map_flags & BPF_F_LINK) {
874 err = 0;
875 /* Let bpf_link handle registration & unregistration.
876 *
877 * Pair with smp_load_acquire() during lookup_elem().
878 */
879 smp_store_release(&kvalue->common.state, BPF_STRUCT_OPS_STATE_READY);
880 goto unlock;
881 }
882
883 err = st_ops->reg(kdata, NULL);
884 if (likely(!err)) {
885 /* This refcnt increment on the map here after
886 * 'st_ops->reg()' is secure since the state of the
887 * map must be set to INIT at this moment, and thus
888 * bpf_struct_ops_map_delete_elem() can't unregister
889 * or transition it to TOBEFREE concurrently.
890 */
891 bpf_map_inc(map);
892 /* Pair with smp_load_acquire() during lookup_elem().
893 * It ensures the above udata updates (e.g. prog->aux->id)
894 * can be seen once BPF_STRUCT_OPS_STATE_INUSE is set.
895 */
896 smp_store_release(&kvalue->common.state, BPF_STRUCT_OPS_STATE_INUSE);
897 goto unlock;
898 }
899
900 /* Error during st_ops->reg(). Can happen if this struct_ops needs to be
901 * verified as a whole, after all init_member() calls. Can also happen if
902 * there was a race in registering the struct_ops (under the same name) to
903 * a sub-system through different struct_ops's maps.
904 */
905
906 reset_unlock:
907 bpf_struct_ops_map_free_ksyms(st_map);
908 bpf_struct_ops_map_free_image(st_map);
909 bpf_struct_ops_map_put_progs(st_map);
910 memset(uvalue, 0, map->value_size);
911 memset(kvalue, 0, map->value_size);
912 unlock:
913 kfree(tlinks);
914 mutex_unlock(&st_map->lock);
915 if (!err)
916 bpf_struct_ops_map_add_ksyms(st_map);
917 return err;
918 }
919
bpf_struct_ops_map_delete_elem(struct bpf_map * map,void * key)920 static long bpf_struct_ops_map_delete_elem(struct bpf_map *map, void *key)
921 {
922 enum bpf_struct_ops_state prev_state;
923 struct bpf_struct_ops_map *st_map;
924
925 st_map = (struct bpf_struct_ops_map *)map;
926 if (st_map->map.map_flags & BPF_F_LINK)
927 return -EOPNOTSUPP;
928
929 prev_state = cmpxchg(&st_map->kvalue.common.state,
930 BPF_STRUCT_OPS_STATE_INUSE,
931 BPF_STRUCT_OPS_STATE_TOBEFREE);
932 switch (prev_state) {
933 case BPF_STRUCT_OPS_STATE_INUSE:
934 st_map->st_ops_desc->st_ops->unreg(&st_map->kvalue.data, NULL);
935 bpf_map_put(map);
936 return 0;
937 case BPF_STRUCT_OPS_STATE_TOBEFREE:
938 return -EINPROGRESS;
939 case BPF_STRUCT_OPS_STATE_INIT:
940 return -ENOENT;
941 default:
942 WARN_ON_ONCE(1);
943 /* Should never happen. Treat it as not found. */
944 return -ENOENT;
945 }
946 }
947
bpf_struct_ops_map_seq_show_elem(struct bpf_map * map,void * key,struct seq_file * m)948 static void bpf_struct_ops_map_seq_show_elem(struct bpf_map *map, void *key,
949 struct seq_file *m)
950 {
951 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
952 void *value;
953 int err;
954
955 value = kmalloc(map->value_size, GFP_USER | __GFP_NOWARN);
956 if (!value)
957 return;
958
959 err = bpf_struct_ops_map_sys_lookup_elem(map, key, value);
960 if (!err) {
961 btf_type_seq_show(st_map->btf,
962 map->btf_vmlinux_value_type_id,
963 value, m);
964 seq_putc(m, '\n');
965 }
966
967 kfree(value);
968 }
969
__bpf_struct_ops_map_free(struct bpf_map * map)970 static void __bpf_struct_ops_map_free(struct bpf_map *map)
971 {
972 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
973
974 if (st_map->links)
975 bpf_struct_ops_map_put_progs(st_map);
976 if (st_map->ksyms)
977 bpf_struct_ops_map_free_ksyms(st_map);
978 bpf_map_area_free(st_map->links);
979 bpf_map_area_free(st_map->ksyms);
980 bpf_struct_ops_map_free_image(st_map);
981 bpf_map_area_free(st_map->uvalue);
982 bpf_map_area_free(st_map);
983 }
984
bpf_struct_ops_map_free(struct bpf_map * map)985 static void bpf_struct_ops_map_free(struct bpf_map *map)
986 {
987 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
988
989 /* st_ops->owner was acquired during map_alloc to implicitly holds
990 * the btf's refcnt. The acquire was only done when btf_is_module()
991 * st_map->btf cannot be NULL here.
992 */
993 if (btf_is_module(st_map->btf))
994 module_put(st_map->st_ops_desc->st_ops->owner);
995
996 bpf_struct_ops_map_dissoc_progs(st_map);
997
998 bpf_struct_ops_map_del_ksyms(st_map);
999
1000 /* The struct_ops's function may switch to another struct_ops.
1001 *
1002 * For example, bpf_tcp_cc_x->init() may switch to
1003 * another tcp_cc_y by calling
1004 * setsockopt(TCP_CONGESTION, "tcp_cc_y").
1005 * During the switch, bpf_struct_ops_put(tcp_cc_x) is called
1006 * and its refcount may reach 0 which then free its
1007 * trampoline image while tcp_cc_x is still running.
1008 *
1009 * A vanilla rcu gp is to wait for all bpf-tcp-cc prog
1010 * to finish. bpf-tcp-cc prog is non sleepable.
1011 * A rcu_tasks gp is to wait for the last few insn
1012 * in the tramopline image to finish before releasing
1013 * the trampoline image.
1014 */
1015 synchronize_rcu_mult(call_rcu, call_rcu_tasks);
1016
1017 __bpf_struct_ops_map_free(map);
1018 }
1019
bpf_struct_ops_map_alloc_check(union bpf_attr * attr)1020 static int bpf_struct_ops_map_alloc_check(union bpf_attr *attr)
1021 {
1022 if (attr->key_size != sizeof(unsigned int) || attr->max_entries != 1 ||
1023 (attr->map_flags & ~(BPF_F_LINK | BPF_F_VTYPE_BTF_OBJ_FD)) ||
1024 !attr->btf_vmlinux_value_type_id)
1025 return -EINVAL;
1026 return 0;
1027 }
1028
count_func_ptrs(const struct btf * btf,const struct btf_type * t)1029 static u32 count_func_ptrs(const struct btf *btf, const struct btf_type *t)
1030 {
1031 int i;
1032 u32 count;
1033 const struct btf_member *member;
1034
1035 count = 0;
1036 for_each_member(i, t, member)
1037 if (btf_type_resolve_func_ptr(btf, member->type, NULL))
1038 count++;
1039 return count;
1040 }
1041
bpf_struct_ops_map_alloc(union bpf_attr * attr)1042 static struct bpf_map *bpf_struct_ops_map_alloc(union bpf_attr *attr)
1043 {
1044 const struct bpf_struct_ops_desc *st_ops_desc;
1045 size_t st_map_size;
1046 struct bpf_struct_ops_map *st_map;
1047 const struct btf_type *t, *vt;
1048 struct module *mod = NULL;
1049 struct bpf_map *map;
1050 struct btf *btf;
1051 int ret;
1052
1053 if (attr->map_flags & BPF_F_VTYPE_BTF_OBJ_FD) {
1054 /* The map holds btf for its whole life time. */
1055 btf = btf_get_by_fd(attr->value_type_btf_obj_fd);
1056 if (IS_ERR(btf))
1057 return ERR_CAST(btf);
1058 if (!btf_is_module(btf)) {
1059 btf_put(btf);
1060 return ERR_PTR(-EINVAL);
1061 }
1062
1063 mod = btf_try_get_module(btf);
1064 /* mod holds a refcnt to btf. We don't need an extra refcnt
1065 * here.
1066 */
1067 btf_put(btf);
1068 if (!mod)
1069 return ERR_PTR(-EINVAL);
1070 } else {
1071 btf = bpf_get_btf_vmlinux();
1072 if (IS_ERR(btf))
1073 return ERR_CAST(btf);
1074 if (!btf)
1075 return ERR_PTR(-ENOTSUPP);
1076 }
1077
1078 st_ops_desc = bpf_struct_ops_find_value(btf, attr->btf_vmlinux_value_type_id);
1079 if (!st_ops_desc) {
1080 ret = -ENOTSUPP;
1081 goto errout;
1082 }
1083
1084 vt = st_ops_desc->value_type;
1085 if (attr->value_size != vt->size) {
1086 ret = -EINVAL;
1087 goto errout;
1088 }
1089
1090 t = st_ops_desc->type;
1091
1092 st_map_size = sizeof(*st_map) +
1093 /* kvalue stores the
1094 * struct bpf_struct_ops_tcp_congestions_ops
1095 */
1096 (vt->size - sizeof(struct bpf_struct_ops_value));
1097
1098 st_map = bpf_map_area_alloc(st_map_size, NUMA_NO_NODE);
1099 if (!st_map) {
1100 ret = -ENOMEM;
1101 goto errout;
1102 }
1103
1104 st_map->st_ops_desc = st_ops_desc;
1105 map = &st_map->map;
1106
1107 st_map->uvalue = bpf_map_area_alloc(vt->size, NUMA_NO_NODE);
1108 st_map->funcs_cnt = count_func_ptrs(btf, t);
1109 st_map->links =
1110 bpf_map_area_alloc(st_map->funcs_cnt * sizeof(struct bpf_link *),
1111 NUMA_NO_NODE);
1112
1113 st_map->ksyms =
1114 bpf_map_area_alloc(st_map->funcs_cnt * sizeof(struct bpf_ksym *),
1115 NUMA_NO_NODE);
1116 if (!st_map->uvalue || !st_map->links || !st_map->ksyms) {
1117 ret = -ENOMEM;
1118 goto errout_free;
1119 }
1120 st_map->btf = btf;
1121
1122 mutex_init(&st_map->lock);
1123 bpf_map_init_from_attr(map, attr);
1124
1125 return map;
1126
1127 errout_free:
1128 __bpf_struct_ops_map_free(map);
1129 errout:
1130 module_put(mod);
1131
1132 return ERR_PTR(ret);
1133 }
1134
bpf_struct_ops_map_mem_usage(const struct bpf_map * map)1135 static u64 bpf_struct_ops_map_mem_usage(const struct bpf_map *map)
1136 {
1137 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
1138 const struct bpf_struct_ops_desc *st_ops_desc = st_map->st_ops_desc;
1139 const struct btf_type *vt = st_ops_desc->value_type;
1140 u64 usage;
1141
1142 usage = sizeof(*st_map) +
1143 vt->size - sizeof(struct bpf_struct_ops_value);
1144 usage += vt->size;
1145 usage += st_map->funcs_cnt * sizeof(struct bpf_link *);
1146 usage += st_map->funcs_cnt * sizeof(struct bpf_ksym *);
1147 usage += PAGE_SIZE;
1148 return usage;
1149 }
1150
1151 BTF_ID_LIST_SINGLE(bpf_struct_ops_map_btf_ids, struct, bpf_struct_ops_map)
1152 const struct bpf_map_ops bpf_struct_ops_map_ops = {
1153 .map_alloc_check = bpf_struct_ops_map_alloc_check,
1154 .map_alloc = bpf_struct_ops_map_alloc,
1155 .map_free = bpf_struct_ops_map_free,
1156 .map_get_next_key = bpf_struct_ops_map_get_next_key,
1157 .map_lookup_elem = bpf_struct_ops_map_lookup_elem,
1158 .map_delete_elem = bpf_struct_ops_map_delete_elem,
1159 .map_update_elem = bpf_struct_ops_map_update_elem,
1160 .map_seq_show_elem = bpf_struct_ops_map_seq_show_elem,
1161 .map_mem_usage = bpf_struct_ops_map_mem_usage,
1162 .map_btf_id = &bpf_struct_ops_map_btf_ids[0],
1163 };
1164
1165 /* "const void *" because some subsystem is
1166 * passing a const (e.g. const struct tcp_congestion_ops *)
1167 */
bpf_struct_ops_get(const void * kdata)1168 bool bpf_struct_ops_get(const void *kdata)
1169 {
1170 struct bpf_struct_ops_value *kvalue;
1171 struct bpf_struct_ops_map *st_map;
1172 struct bpf_map *map;
1173
1174 kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
1175 st_map = container_of(kvalue, struct bpf_struct_ops_map, kvalue);
1176
1177 map = __bpf_map_inc_not_zero(&st_map->map, false);
1178 return !IS_ERR(map);
1179 }
1180 EXPORT_SYMBOL_GPL(bpf_struct_ops_get);
1181
bpf_struct_ops_put(const void * kdata)1182 void bpf_struct_ops_put(const void *kdata)
1183 {
1184 struct bpf_struct_ops_value *kvalue;
1185 struct bpf_struct_ops_map *st_map;
1186
1187 kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
1188 st_map = container_of(kvalue, struct bpf_struct_ops_map, kvalue);
1189
1190 bpf_map_put(&st_map->map);
1191 }
1192 EXPORT_SYMBOL_GPL(bpf_struct_ops_put);
1193
bpf_struct_ops_id(const void * kdata)1194 u32 bpf_struct_ops_id(const void *kdata)
1195 {
1196 struct bpf_struct_ops_value *kvalue;
1197 struct bpf_struct_ops_map *st_map;
1198
1199 kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
1200 st_map = container_of(kvalue, struct bpf_struct_ops_map, kvalue);
1201
1202 return st_map->map.id;
1203 }
1204 EXPORT_SYMBOL_GPL(bpf_struct_ops_id);
1205
bpf_struct_ops_valid_to_reg(struct bpf_map * map)1206 static bool bpf_struct_ops_valid_to_reg(struct bpf_map *map)
1207 {
1208 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
1209
1210 return map->map_type == BPF_MAP_TYPE_STRUCT_OPS &&
1211 map->map_flags & BPF_F_LINK &&
1212 /* Pair with smp_store_release() during map_update */
1213 smp_load_acquire(&st_map->kvalue.common.state) == BPF_STRUCT_OPS_STATE_READY;
1214 }
1215
bpf_struct_ops_map_link_dealloc(struct bpf_link * link)1216 static void bpf_struct_ops_map_link_dealloc(struct bpf_link *link)
1217 {
1218 struct bpf_struct_ops_link *st_link;
1219 struct bpf_struct_ops_map *st_map;
1220
1221 st_link = container_of(link, struct bpf_struct_ops_link, link);
1222 st_map = (struct bpf_struct_ops_map *)
1223 rcu_dereference_protected(st_link->map, true);
1224 if (st_map) {
1225 st_map->st_ops_desc->st_ops->unreg(&st_map->kvalue.data, link);
1226 bpf_map_put(&st_map->map);
1227 }
1228 kfree(st_link);
1229 }
1230
bpf_struct_ops_map_link_show_fdinfo(const struct bpf_link * link,struct seq_file * seq)1231 static void bpf_struct_ops_map_link_show_fdinfo(const struct bpf_link *link,
1232 struct seq_file *seq)
1233 {
1234 struct bpf_struct_ops_link *st_link;
1235 struct bpf_map *map;
1236
1237 st_link = container_of(link, struct bpf_struct_ops_link, link);
1238 rcu_read_lock();
1239 map = rcu_dereference(st_link->map);
1240 if (map)
1241 seq_printf(seq, "map_id:\t%d\n", map->id);
1242 rcu_read_unlock();
1243 }
1244
bpf_struct_ops_map_link_fill_link_info(const struct bpf_link * link,struct bpf_link_info * info)1245 static int bpf_struct_ops_map_link_fill_link_info(const struct bpf_link *link,
1246 struct bpf_link_info *info)
1247 {
1248 struct bpf_struct_ops_link *st_link;
1249 struct bpf_map *map;
1250
1251 st_link = container_of(link, struct bpf_struct_ops_link, link);
1252 rcu_read_lock();
1253 map = rcu_dereference(st_link->map);
1254 if (map)
1255 info->struct_ops.map_id = map->id;
1256 rcu_read_unlock();
1257 return 0;
1258 }
1259
bpf_struct_ops_map_link_update(struct bpf_link * link,struct bpf_map * new_map,struct bpf_map * expected_old_map)1260 static int bpf_struct_ops_map_link_update(struct bpf_link *link, struct bpf_map *new_map,
1261 struct bpf_map *expected_old_map)
1262 {
1263 struct bpf_struct_ops_map *st_map, *old_st_map;
1264 struct bpf_map *old_map;
1265 struct bpf_struct_ops_link *st_link;
1266 int err;
1267
1268 st_link = container_of(link, struct bpf_struct_ops_link, link);
1269 st_map = container_of(new_map, struct bpf_struct_ops_map, map);
1270
1271 if (!bpf_struct_ops_valid_to_reg(new_map))
1272 return -EINVAL;
1273
1274 if (!st_map->st_ops_desc->st_ops->update)
1275 return -EOPNOTSUPP;
1276
1277 mutex_lock(&update_mutex);
1278
1279 old_map = rcu_dereference_protected(st_link->map, lockdep_is_held(&update_mutex));
1280 if (!old_map) {
1281 err = -ENOLINK;
1282 goto err_out;
1283 }
1284 if (expected_old_map && old_map != expected_old_map) {
1285 err = -EPERM;
1286 goto err_out;
1287 }
1288
1289 old_st_map = container_of(old_map, struct bpf_struct_ops_map, map);
1290 /* The new and old struct_ops must be the same type. */
1291 if (st_map->st_ops_desc != old_st_map->st_ops_desc) {
1292 err = -EINVAL;
1293 goto err_out;
1294 }
1295
1296 err = st_map->st_ops_desc->st_ops->update(st_map->kvalue.data, old_st_map->kvalue.data, link);
1297 if (err)
1298 goto err_out;
1299
1300 bpf_map_inc(new_map);
1301 rcu_assign_pointer(st_link->map, new_map);
1302 bpf_map_put(old_map);
1303
1304 err_out:
1305 mutex_unlock(&update_mutex);
1306
1307 return err;
1308 }
1309
bpf_struct_ops_map_link_detach(struct bpf_link * link)1310 static int bpf_struct_ops_map_link_detach(struct bpf_link *link)
1311 {
1312 struct bpf_struct_ops_link *st_link = container_of(link, struct bpf_struct_ops_link, link);
1313 struct bpf_struct_ops_map *st_map;
1314 struct bpf_map *map;
1315
1316 mutex_lock(&update_mutex);
1317
1318 map = rcu_dereference_protected(st_link->map, lockdep_is_held(&update_mutex));
1319 if (!map) {
1320 mutex_unlock(&update_mutex);
1321 return 0;
1322 }
1323 st_map = container_of(map, struct bpf_struct_ops_map, map);
1324
1325 st_map->st_ops_desc->st_ops->unreg(&st_map->kvalue.data, link);
1326
1327 RCU_INIT_POINTER(st_link->map, NULL);
1328 /* Pair with bpf_map_get() in bpf_struct_ops_link_create() or
1329 * bpf_map_inc() in bpf_struct_ops_map_link_update().
1330 */
1331 bpf_map_put(&st_map->map);
1332
1333 mutex_unlock(&update_mutex);
1334
1335 wake_up_interruptible_poll(&st_link->wait_hup, EPOLLHUP);
1336
1337 return 0;
1338 }
1339
bpf_struct_ops_map_link_poll(struct file * file,struct poll_table_struct * pts)1340 static __poll_t bpf_struct_ops_map_link_poll(struct file *file,
1341 struct poll_table_struct *pts)
1342 {
1343 struct bpf_struct_ops_link *st_link = file->private_data;
1344
1345 poll_wait(file, &st_link->wait_hup, pts);
1346
1347 return rcu_access_pointer(st_link->map) ? 0 : EPOLLHUP;
1348 }
1349
1350 static const struct bpf_link_ops bpf_struct_ops_map_lops = {
1351 .dealloc = bpf_struct_ops_map_link_dealloc,
1352 .detach = bpf_struct_ops_map_link_detach,
1353 .show_fdinfo = bpf_struct_ops_map_link_show_fdinfo,
1354 .fill_link_info = bpf_struct_ops_map_link_fill_link_info,
1355 .update_map = bpf_struct_ops_map_link_update,
1356 .poll = bpf_struct_ops_map_link_poll,
1357 };
1358
bpf_struct_ops_link_create(union bpf_attr * attr)1359 int bpf_struct_ops_link_create(union bpf_attr *attr)
1360 {
1361 struct bpf_struct_ops_link *link = NULL;
1362 struct bpf_link_primer link_primer;
1363 struct bpf_struct_ops_map *st_map;
1364 struct bpf_map *map;
1365 int err;
1366
1367 map = bpf_map_get(attr->link_create.map_fd);
1368 if (IS_ERR(map))
1369 return PTR_ERR(map);
1370
1371 st_map = (struct bpf_struct_ops_map *)map;
1372
1373 if (!bpf_struct_ops_valid_to_reg(map)) {
1374 err = -EINVAL;
1375 goto err_out;
1376 }
1377
1378 link = kzalloc_obj(*link, GFP_USER);
1379 if (!link) {
1380 err = -ENOMEM;
1381 goto err_out;
1382 }
1383 bpf_link_init(&link->link, BPF_LINK_TYPE_STRUCT_OPS, &bpf_struct_ops_map_lops, NULL,
1384 attr->link_create.attach_type);
1385
1386 err = bpf_link_prime(&link->link, &link_primer);
1387 if (err)
1388 goto err_out;
1389
1390 init_waitqueue_head(&link->wait_hup);
1391
1392 /* Hold the update_mutex such that the subsystem cannot
1393 * do link->ops->detach() before the link is fully initialized.
1394 */
1395 mutex_lock(&update_mutex);
1396 err = st_map->st_ops_desc->st_ops->reg(st_map->kvalue.data, &link->link);
1397 if (err) {
1398 mutex_unlock(&update_mutex);
1399 bpf_link_cleanup(&link_primer);
1400 link = NULL;
1401 goto err_out;
1402 }
1403 RCU_INIT_POINTER(link->map, map);
1404 mutex_unlock(&update_mutex);
1405
1406 return bpf_link_settle(&link_primer);
1407
1408 err_out:
1409 bpf_map_put(map);
1410 kfree(link);
1411 return err;
1412 }
1413
bpf_prog_assoc_struct_ops(struct bpf_prog * prog,struct bpf_map * map)1414 int bpf_prog_assoc_struct_ops(struct bpf_prog *prog, struct bpf_map *map)
1415 {
1416 struct bpf_map *st_ops_assoc;
1417
1418 guard(mutex)(&prog->aux->st_ops_assoc_mutex);
1419
1420 st_ops_assoc = rcu_dereference_protected(prog->aux->st_ops_assoc,
1421 lockdep_is_held(&prog->aux->st_ops_assoc_mutex));
1422 if (st_ops_assoc && st_ops_assoc == map)
1423 return 0;
1424
1425 if (st_ops_assoc) {
1426 if (prog->type != BPF_PROG_TYPE_STRUCT_OPS)
1427 return -EBUSY;
1428
1429 rcu_assign_pointer(prog->aux->st_ops_assoc, BPF_PTR_POISON);
1430 } else {
1431 /*
1432 * struct_ops map does not track associated non-struct_ops programs.
1433 * Bump the refcount to make sure st_ops_assoc is always valid.
1434 */
1435 if (prog->type != BPF_PROG_TYPE_STRUCT_OPS)
1436 bpf_map_inc(map);
1437
1438 rcu_assign_pointer(prog->aux->st_ops_assoc, map);
1439 }
1440
1441 return 0;
1442 }
1443
bpf_prog_disassoc_struct_ops(struct bpf_prog * prog)1444 void bpf_prog_disassoc_struct_ops(struct bpf_prog *prog)
1445 {
1446 struct bpf_map *st_ops_assoc;
1447
1448 guard(mutex)(&prog->aux->st_ops_assoc_mutex);
1449
1450 st_ops_assoc = rcu_dereference_protected(prog->aux->st_ops_assoc,
1451 lockdep_is_held(&prog->aux->st_ops_assoc_mutex));
1452 if (!st_ops_assoc || st_ops_assoc == BPF_PTR_POISON)
1453 return;
1454
1455 if (prog->type != BPF_PROG_TYPE_STRUCT_OPS)
1456 bpf_map_put(st_ops_assoc);
1457
1458 RCU_INIT_POINTER(prog->aux->st_ops_assoc, NULL);
1459 }
1460
1461 /*
1462 * Get a reference to the struct_ops struct (i.e., kdata) associated with a
1463 * program. Should only be called in BPF program context (e.g., in a kfunc).
1464 *
1465 * If the returned pointer is not NULL, it must points to a valid struct_ops.
1466 * The struct_ops map is not guaranteed to be initialized nor attached.
1467 * Kernel struct_ops implementers are responsible for tracking and checking
1468 * the state of the struct_ops if the use case requires an initialized or
1469 * attached struct_ops.
1470 */
bpf_prog_get_assoc_struct_ops(const struct bpf_prog_aux * aux)1471 void *bpf_prog_get_assoc_struct_ops(const struct bpf_prog_aux *aux)
1472 {
1473 struct bpf_struct_ops_map *st_map;
1474 struct bpf_map *st_ops_assoc;
1475
1476 st_ops_assoc = rcu_dereference_check(aux->st_ops_assoc, bpf_rcu_lock_held());
1477 if (!st_ops_assoc || st_ops_assoc == BPF_PTR_POISON)
1478 return NULL;
1479
1480 st_map = (struct bpf_struct_ops_map *)st_ops_assoc;
1481
1482 return &st_map->kvalue.data;
1483 }
1484 EXPORT_SYMBOL_GPL(bpf_prog_get_assoc_struct_ops);
1485
bpf_map_struct_ops_info_fill(struct bpf_map_info * info,struct bpf_map * map)1486 void bpf_map_struct_ops_info_fill(struct bpf_map_info *info, struct bpf_map *map)
1487 {
1488 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
1489
1490 info->btf_vmlinux_id = btf_obj_id(st_map->btf);
1491 }
1492