xref: /linux/kernel/bpf/bpf_struct_ops.c (revision 06ce23ad57c8e378b86ef3f439b2e08bcb5d05eb)
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 
15 enum bpf_struct_ops_state {
16 	BPF_STRUCT_OPS_STATE_INIT,
17 	BPF_STRUCT_OPS_STATE_INUSE,
18 	BPF_STRUCT_OPS_STATE_TOBEFREE,
19 };
20 
21 #define BPF_STRUCT_OPS_COMMON_VALUE			\
22 	refcount_t refcnt;				\
23 	enum bpf_struct_ops_state state
24 
25 struct bpf_struct_ops_value {
26 	BPF_STRUCT_OPS_COMMON_VALUE;
27 	char data[] ____cacheline_aligned_in_smp;
28 };
29 
30 struct bpf_struct_ops_map {
31 	struct bpf_map map;
32 	struct rcu_head rcu;
33 	const struct bpf_struct_ops *st_ops;
34 	/* protect map_update */
35 	struct mutex lock;
36 	/* link has all the bpf_links that is populated
37 	 * to the func ptr of the kernel's struct
38 	 * (in kvalue.data).
39 	 */
40 	struct bpf_link **links;
41 	/* image is a page that has all the trampolines
42 	 * that stores the func args before calling the bpf_prog.
43 	 * A PAGE_SIZE "image" is enough to store all trampoline for
44 	 * "links[]".
45 	 */
46 	void *image;
47 	/* uvalue->data stores the kernel struct
48 	 * (e.g. tcp_congestion_ops) that is more useful
49 	 * to userspace than the kvalue.  For example,
50 	 * the bpf_prog's id is stored instead of the kernel
51 	 * address of a func ptr.
52 	 */
53 	struct bpf_struct_ops_value *uvalue;
54 	/* kvalue.data stores the actual kernel's struct
55 	 * (e.g. tcp_congestion_ops) that will be
56 	 * registered to the kernel subsystem.
57 	 */
58 	struct bpf_struct_ops_value kvalue;
59 };
60 
61 #define VALUE_PREFIX "bpf_struct_ops_"
62 #define VALUE_PREFIX_LEN (sizeof(VALUE_PREFIX) - 1)
63 
64 /* bpf_struct_ops_##_name (e.g. bpf_struct_ops_tcp_congestion_ops) is
65  * the map's value exposed to the userspace and its btf-type-id is
66  * stored at the map->btf_vmlinux_value_type_id.
67  *
68  */
69 #define BPF_STRUCT_OPS_TYPE(_name)				\
70 extern struct bpf_struct_ops bpf_##_name;			\
71 								\
72 struct bpf_struct_ops_##_name {						\
73 	BPF_STRUCT_OPS_COMMON_VALUE;				\
74 	struct _name data ____cacheline_aligned_in_smp;		\
75 };
76 #include "bpf_struct_ops_types.h"
77 #undef BPF_STRUCT_OPS_TYPE
78 
79 enum {
80 #define BPF_STRUCT_OPS_TYPE(_name) BPF_STRUCT_OPS_TYPE_##_name,
81 #include "bpf_struct_ops_types.h"
82 #undef BPF_STRUCT_OPS_TYPE
83 	__NR_BPF_STRUCT_OPS_TYPE,
84 };
85 
86 static struct bpf_struct_ops * const bpf_struct_ops[] = {
87 #define BPF_STRUCT_OPS_TYPE(_name)				\
88 	[BPF_STRUCT_OPS_TYPE_##_name] = &bpf_##_name,
89 #include "bpf_struct_ops_types.h"
90 #undef BPF_STRUCT_OPS_TYPE
91 };
92 
93 const struct bpf_verifier_ops bpf_struct_ops_verifier_ops = {
94 };
95 
96 const struct bpf_prog_ops bpf_struct_ops_prog_ops = {
97 #ifdef CONFIG_NET
98 	.test_run = bpf_struct_ops_test_run,
99 #endif
100 };
101 
102 static const struct btf_type *module_type;
103 
104 void bpf_struct_ops_init(struct btf *btf, struct bpf_verifier_log *log)
105 {
106 	s32 type_id, value_id, module_id;
107 	const struct btf_member *member;
108 	struct bpf_struct_ops *st_ops;
109 	const struct btf_type *t;
110 	char value_name[128];
111 	const char *mname;
112 	u32 i, j;
113 
114 	/* Ensure BTF type is emitted for "struct bpf_struct_ops_##_name" */
115 #define BPF_STRUCT_OPS_TYPE(_name) BTF_TYPE_EMIT(struct bpf_struct_ops_##_name);
116 #include "bpf_struct_ops_types.h"
117 #undef BPF_STRUCT_OPS_TYPE
118 
119 	module_id = btf_find_by_name_kind(btf, "module", BTF_KIND_STRUCT);
120 	if (module_id < 0) {
121 		pr_warn("Cannot find struct module in btf_vmlinux\n");
122 		return;
123 	}
124 	module_type = btf_type_by_id(btf, module_id);
125 
126 	for (i = 0; i < ARRAY_SIZE(bpf_struct_ops); i++) {
127 		st_ops = bpf_struct_ops[i];
128 
129 		if (strlen(st_ops->name) + VALUE_PREFIX_LEN >=
130 		    sizeof(value_name)) {
131 			pr_warn("struct_ops name %s is too long\n",
132 				st_ops->name);
133 			continue;
134 		}
135 		sprintf(value_name, "%s%s", VALUE_PREFIX, st_ops->name);
136 
137 		value_id = btf_find_by_name_kind(btf, value_name,
138 						 BTF_KIND_STRUCT);
139 		if (value_id < 0) {
140 			pr_warn("Cannot find struct %s in btf_vmlinux\n",
141 				value_name);
142 			continue;
143 		}
144 
145 		type_id = btf_find_by_name_kind(btf, st_ops->name,
146 						BTF_KIND_STRUCT);
147 		if (type_id < 0) {
148 			pr_warn("Cannot find struct %s in btf_vmlinux\n",
149 				st_ops->name);
150 			continue;
151 		}
152 		t = btf_type_by_id(btf, type_id);
153 		if (btf_type_vlen(t) > BPF_STRUCT_OPS_MAX_NR_MEMBERS) {
154 			pr_warn("Cannot support #%u members in struct %s\n",
155 				btf_type_vlen(t), st_ops->name);
156 			continue;
157 		}
158 
159 		for_each_member(j, t, member) {
160 			const struct btf_type *func_proto;
161 
162 			mname = btf_name_by_offset(btf, member->name_off);
163 			if (!*mname) {
164 				pr_warn("anon member in struct %s is not supported\n",
165 					st_ops->name);
166 				break;
167 			}
168 
169 			if (__btf_member_bitfield_size(t, member)) {
170 				pr_warn("bit field member %s in struct %s is not supported\n",
171 					mname, st_ops->name);
172 				break;
173 			}
174 
175 			func_proto = btf_type_resolve_func_ptr(btf,
176 							       member->type,
177 							       NULL);
178 			if (func_proto &&
179 			    btf_distill_func_proto(log, btf,
180 						   func_proto, mname,
181 						   &st_ops->func_models[j])) {
182 				pr_warn("Error in parsing func ptr %s in struct %s\n",
183 					mname, st_ops->name);
184 				break;
185 			}
186 		}
187 
188 		if (j == btf_type_vlen(t)) {
189 			if (st_ops->init(btf)) {
190 				pr_warn("Error in init bpf_struct_ops %s\n",
191 					st_ops->name);
192 			} else {
193 				st_ops->type_id = type_id;
194 				st_ops->type = t;
195 				st_ops->value_id = value_id;
196 				st_ops->value_type = btf_type_by_id(btf,
197 								    value_id);
198 			}
199 		}
200 	}
201 }
202 
203 extern struct btf *btf_vmlinux;
204 
205 static const struct bpf_struct_ops *
206 bpf_struct_ops_find_value(u32 value_id)
207 {
208 	unsigned int i;
209 
210 	if (!value_id || !btf_vmlinux)
211 		return NULL;
212 
213 	for (i = 0; i < ARRAY_SIZE(bpf_struct_ops); i++) {
214 		if (bpf_struct_ops[i]->value_id == value_id)
215 			return bpf_struct_ops[i];
216 	}
217 
218 	return NULL;
219 }
220 
221 const struct bpf_struct_ops *bpf_struct_ops_find(u32 type_id)
222 {
223 	unsigned int i;
224 
225 	if (!type_id || !btf_vmlinux)
226 		return NULL;
227 
228 	for (i = 0; i < ARRAY_SIZE(bpf_struct_ops); i++) {
229 		if (bpf_struct_ops[i]->type_id == type_id)
230 			return bpf_struct_ops[i];
231 	}
232 
233 	return NULL;
234 }
235 
236 static int bpf_struct_ops_map_get_next_key(struct bpf_map *map, void *key,
237 					   void *next_key)
238 {
239 	if (key && *(u32 *)key == 0)
240 		return -ENOENT;
241 
242 	*(u32 *)next_key = 0;
243 	return 0;
244 }
245 
246 int bpf_struct_ops_map_sys_lookup_elem(struct bpf_map *map, void *key,
247 				       void *value)
248 {
249 	struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
250 	struct bpf_struct_ops_value *uvalue, *kvalue;
251 	enum bpf_struct_ops_state state;
252 
253 	if (unlikely(*(u32 *)key != 0))
254 		return -ENOENT;
255 
256 	kvalue = &st_map->kvalue;
257 	/* Pair with smp_store_release() during map_update */
258 	state = smp_load_acquire(&kvalue->state);
259 	if (state == BPF_STRUCT_OPS_STATE_INIT) {
260 		memset(value, 0, map->value_size);
261 		return 0;
262 	}
263 
264 	/* No lock is needed.  state and refcnt do not need
265 	 * to be updated together under atomic context.
266 	 */
267 	uvalue = value;
268 	memcpy(uvalue, st_map->uvalue, map->value_size);
269 	uvalue->state = state;
270 	refcount_set(&uvalue->refcnt, refcount_read(&kvalue->refcnt));
271 
272 	return 0;
273 }
274 
275 static void *bpf_struct_ops_map_lookup_elem(struct bpf_map *map, void *key)
276 {
277 	return ERR_PTR(-EINVAL);
278 }
279 
280 static void bpf_struct_ops_map_put_progs(struct bpf_struct_ops_map *st_map)
281 {
282 	const struct btf_type *t = st_map->st_ops->type;
283 	u32 i;
284 
285 	for (i = 0; i < btf_type_vlen(t); i++) {
286 		if (st_map->links[i]) {
287 			bpf_link_put(st_map->links[i]);
288 			st_map->links[i] = NULL;
289 		}
290 	}
291 }
292 
293 static int check_zero_holes(const struct btf_type *t, void *data)
294 {
295 	const struct btf_member *member;
296 	u32 i, moff, msize, prev_mend = 0;
297 	const struct btf_type *mtype;
298 
299 	for_each_member(i, t, member) {
300 		moff = __btf_member_bit_offset(t, member) / 8;
301 		if (moff > prev_mend &&
302 		    memchr_inv(data + prev_mend, 0, moff - prev_mend))
303 			return -EINVAL;
304 
305 		mtype = btf_type_by_id(btf_vmlinux, member->type);
306 		mtype = btf_resolve_size(btf_vmlinux, mtype, &msize);
307 		if (IS_ERR(mtype))
308 			return PTR_ERR(mtype);
309 		prev_mend = moff + msize;
310 	}
311 
312 	if (t->size > prev_mend &&
313 	    memchr_inv(data + prev_mend, 0, t->size - prev_mend))
314 		return -EINVAL;
315 
316 	return 0;
317 }
318 
319 static void bpf_struct_ops_link_release(struct bpf_link *link)
320 {
321 }
322 
323 static void bpf_struct_ops_link_dealloc(struct bpf_link *link)
324 {
325 	struct bpf_tramp_link *tlink = container_of(link, struct bpf_tramp_link, link);
326 
327 	kfree(tlink);
328 }
329 
330 const struct bpf_link_ops bpf_struct_ops_link_lops = {
331 	.release = bpf_struct_ops_link_release,
332 	.dealloc = bpf_struct_ops_link_dealloc,
333 };
334 
335 int bpf_struct_ops_prepare_trampoline(struct bpf_tramp_links *tlinks,
336 				      struct bpf_tramp_link *link,
337 				      const struct btf_func_model *model,
338 				      void *image, void *image_end)
339 {
340 	u32 flags;
341 
342 	tlinks[BPF_TRAMP_FENTRY].links[0] = link;
343 	tlinks[BPF_TRAMP_FENTRY].nr_links = 1;
344 	/* BPF_TRAMP_F_RET_FENTRY_RET is only used by bpf_struct_ops,
345 	 * and it must be used alone.
346 	 */
347 	flags = model->ret_size > 0 ? BPF_TRAMP_F_RET_FENTRY_RET : 0;
348 	return arch_prepare_bpf_trampoline(NULL, image, image_end,
349 					   model, flags, tlinks, NULL);
350 }
351 
352 static int bpf_struct_ops_map_update_elem(struct bpf_map *map, void *key,
353 					  void *value, u64 flags)
354 {
355 	struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
356 	const struct bpf_struct_ops *st_ops = st_map->st_ops;
357 	struct bpf_struct_ops_value *uvalue, *kvalue;
358 	const struct btf_member *member;
359 	const struct btf_type *t = st_ops->type;
360 	struct bpf_tramp_links *tlinks = NULL;
361 	void *udata, *kdata;
362 	int prog_fd, err = 0;
363 	void *image, *image_end;
364 	u32 i;
365 
366 	if (flags)
367 		return -EINVAL;
368 
369 	if (*(u32 *)key != 0)
370 		return -E2BIG;
371 
372 	err = check_zero_holes(st_ops->value_type, value);
373 	if (err)
374 		return err;
375 
376 	uvalue = value;
377 	err = check_zero_holes(t, uvalue->data);
378 	if (err)
379 		return err;
380 
381 	if (uvalue->state || refcount_read(&uvalue->refcnt))
382 		return -EINVAL;
383 
384 	tlinks = kcalloc(BPF_TRAMP_MAX, sizeof(*tlinks), GFP_KERNEL);
385 	if (!tlinks)
386 		return -ENOMEM;
387 
388 	uvalue = (struct bpf_struct_ops_value *)st_map->uvalue;
389 	kvalue = (struct bpf_struct_ops_value *)&st_map->kvalue;
390 
391 	mutex_lock(&st_map->lock);
392 
393 	if (kvalue->state != BPF_STRUCT_OPS_STATE_INIT) {
394 		err = -EBUSY;
395 		goto unlock;
396 	}
397 
398 	memcpy(uvalue, value, map->value_size);
399 
400 	udata = &uvalue->data;
401 	kdata = &kvalue->data;
402 	image = st_map->image;
403 	image_end = st_map->image + PAGE_SIZE;
404 
405 	for_each_member(i, t, member) {
406 		const struct btf_type *mtype, *ptype;
407 		struct bpf_prog *prog;
408 		struct bpf_tramp_link *link;
409 		u32 moff;
410 
411 		moff = __btf_member_bit_offset(t, member) / 8;
412 		ptype = btf_type_resolve_ptr(btf_vmlinux, member->type, NULL);
413 		if (ptype == module_type) {
414 			if (*(void **)(udata + moff))
415 				goto reset_unlock;
416 			*(void **)(kdata + moff) = BPF_MODULE_OWNER;
417 			continue;
418 		}
419 
420 		err = st_ops->init_member(t, member, kdata, udata);
421 		if (err < 0)
422 			goto reset_unlock;
423 
424 		/* The ->init_member() has handled this member */
425 		if (err > 0)
426 			continue;
427 
428 		/* If st_ops->init_member does not handle it,
429 		 * we will only handle func ptrs and zero-ed members
430 		 * here.  Reject everything else.
431 		 */
432 
433 		/* All non func ptr member must be 0 */
434 		if (!ptype || !btf_type_is_func_proto(ptype)) {
435 			u32 msize;
436 
437 			mtype = btf_type_by_id(btf_vmlinux, member->type);
438 			mtype = btf_resolve_size(btf_vmlinux, mtype, &msize);
439 			if (IS_ERR(mtype)) {
440 				err = PTR_ERR(mtype);
441 				goto reset_unlock;
442 			}
443 
444 			if (memchr_inv(udata + moff, 0, msize)) {
445 				err = -EINVAL;
446 				goto reset_unlock;
447 			}
448 
449 			continue;
450 		}
451 
452 		prog_fd = (int)(*(unsigned long *)(udata + moff));
453 		/* Similar check as the attr->attach_prog_fd */
454 		if (!prog_fd)
455 			continue;
456 
457 		prog = bpf_prog_get(prog_fd);
458 		if (IS_ERR(prog)) {
459 			err = PTR_ERR(prog);
460 			goto reset_unlock;
461 		}
462 
463 		if (prog->type != BPF_PROG_TYPE_STRUCT_OPS ||
464 		    prog->aux->attach_btf_id != st_ops->type_id ||
465 		    prog->expected_attach_type != i) {
466 			bpf_prog_put(prog);
467 			err = -EINVAL;
468 			goto reset_unlock;
469 		}
470 
471 		link = kzalloc(sizeof(*link), GFP_USER);
472 		if (!link) {
473 			bpf_prog_put(prog);
474 			err = -ENOMEM;
475 			goto reset_unlock;
476 		}
477 		bpf_link_init(&link->link, BPF_LINK_TYPE_STRUCT_OPS,
478 			      &bpf_struct_ops_link_lops, prog);
479 		st_map->links[i] = &link->link;
480 
481 		err = bpf_struct_ops_prepare_trampoline(tlinks, link,
482 							&st_ops->func_models[i],
483 							image, image_end);
484 		if (err < 0)
485 			goto reset_unlock;
486 
487 		*(void **)(kdata + moff) = image;
488 		image += err;
489 
490 		/* put prog_id to udata */
491 		*(unsigned long *)(udata + moff) = prog->aux->id;
492 	}
493 
494 	refcount_set(&kvalue->refcnt, 1);
495 	bpf_map_inc(map);
496 
497 	set_memory_rox((long)st_map->image, 1);
498 	err = st_ops->reg(kdata);
499 	if (likely(!err)) {
500 		/* Pair with smp_load_acquire() during lookup_elem().
501 		 * It ensures the above udata updates (e.g. prog->aux->id)
502 		 * can be seen once BPF_STRUCT_OPS_STATE_INUSE is set.
503 		 */
504 		smp_store_release(&kvalue->state, BPF_STRUCT_OPS_STATE_INUSE);
505 		goto unlock;
506 	}
507 
508 	/* Error during st_ops->reg(). Can happen if this struct_ops needs to be
509 	 * verified as a whole, after all init_member() calls. Can also happen if
510 	 * there was a race in registering the struct_ops (under the same name) to
511 	 * a sub-system through different struct_ops's maps.
512 	 */
513 	set_memory_nx((long)st_map->image, 1);
514 	set_memory_rw((long)st_map->image, 1);
515 	bpf_map_put(map);
516 
517 reset_unlock:
518 	bpf_struct_ops_map_put_progs(st_map);
519 	memset(uvalue, 0, map->value_size);
520 	memset(kvalue, 0, map->value_size);
521 unlock:
522 	kfree(tlinks);
523 	mutex_unlock(&st_map->lock);
524 	return err;
525 }
526 
527 static int bpf_struct_ops_map_delete_elem(struct bpf_map *map, void *key)
528 {
529 	enum bpf_struct_ops_state prev_state;
530 	struct bpf_struct_ops_map *st_map;
531 
532 	st_map = (struct bpf_struct_ops_map *)map;
533 	prev_state = cmpxchg(&st_map->kvalue.state,
534 			     BPF_STRUCT_OPS_STATE_INUSE,
535 			     BPF_STRUCT_OPS_STATE_TOBEFREE);
536 	switch (prev_state) {
537 	case BPF_STRUCT_OPS_STATE_INUSE:
538 		st_map->st_ops->unreg(&st_map->kvalue.data);
539 		if (refcount_dec_and_test(&st_map->kvalue.refcnt))
540 			bpf_map_put(map);
541 		return 0;
542 	case BPF_STRUCT_OPS_STATE_TOBEFREE:
543 		return -EINPROGRESS;
544 	case BPF_STRUCT_OPS_STATE_INIT:
545 		return -ENOENT;
546 	default:
547 		WARN_ON_ONCE(1);
548 		/* Should never happen.  Treat it as not found. */
549 		return -ENOENT;
550 	}
551 }
552 
553 static void bpf_struct_ops_map_seq_show_elem(struct bpf_map *map, void *key,
554 					     struct seq_file *m)
555 {
556 	void *value;
557 	int err;
558 
559 	value = kmalloc(map->value_size, GFP_USER | __GFP_NOWARN);
560 	if (!value)
561 		return;
562 
563 	err = bpf_struct_ops_map_sys_lookup_elem(map, key, value);
564 	if (!err) {
565 		btf_type_seq_show(btf_vmlinux, map->btf_vmlinux_value_type_id,
566 				  value, m);
567 		seq_puts(m, "\n");
568 	}
569 
570 	kfree(value);
571 }
572 
573 static void bpf_struct_ops_map_free(struct bpf_map *map)
574 {
575 	struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
576 
577 	if (st_map->links)
578 		bpf_struct_ops_map_put_progs(st_map);
579 	bpf_map_area_free(st_map->links);
580 	bpf_jit_free_exec(st_map->image);
581 	bpf_map_area_free(st_map->uvalue);
582 	bpf_map_area_free(st_map);
583 }
584 
585 static int bpf_struct_ops_map_alloc_check(union bpf_attr *attr)
586 {
587 	if (attr->key_size != sizeof(unsigned int) || attr->max_entries != 1 ||
588 	    attr->map_flags || !attr->btf_vmlinux_value_type_id)
589 		return -EINVAL;
590 	return 0;
591 }
592 
593 static struct bpf_map *bpf_struct_ops_map_alloc(union bpf_attr *attr)
594 {
595 	const struct bpf_struct_ops *st_ops;
596 	size_t st_map_size;
597 	struct bpf_struct_ops_map *st_map;
598 	const struct btf_type *t, *vt;
599 	struct bpf_map *map;
600 
601 	if (!bpf_capable())
602 		return ERR_PTR(-EPERM);
603 
604 	st_ops = bpf_struct_ops_find_value(attr->btf_vmlinux_value_type_id);
605 	if (!st_ops)
606 		return ERR_PTR(-ENOTSUPP);
607 
608 	vt = st_ops->value_type;
609 	if (attr->value_size != vt->size)
610 		return ERR_PTR(-EINVAL);
611 
612 	t = st_ops->type;
613 
614 	st_map_size = sizeof(*st_map) +
615 		/* kvalue stores the
616 		 * struct bpf_struct_ops_tcp_congestions_ops
617 		 */
618 		(vt->size - sizeof(struct bpf_struct_ops_value));
619 
620 	st_map = bpf_map_area_alloc(st_map_size, NUMA_NO_NODE);
621 	if (!st_map)
622 		return ERR_PTR(-ENOMEM);
623 
624 	st_map->st_ops = st_ops;
625 	map = &st_map->map;
626 
627 	st_map->uvalue = bpf_map_area_alloc(vt->size, NUMA_NO_NODE);
628 	st_map->links =
629 		bpf_map_area_alloc(btf_type_vlen(t) * sizeof(struct bpf_links *),
630 				   NUMA_NO_NODE);
631 	st_map->image = bpf_jit_alloc_exec(PAGE_SIZE);
632 	if (!st_map->uvalue || !st_map->links || !st_map->image) {
633 		bpf_struct_ops_map_free(map);
634 		return ERR_PTR(-ENOMEM);
635 	}
636 
637 	mutex_init(&st_map->lock);
638 	set_vm_flush_reset_perms(st_map->image);
639 	bpf_map_init_from_attr(map, attr);
640 
641 	return map;
642 }
643 
644 static u64 bpf_struct_ops_map_mem_usage(const struct bpf_map *map)
645 {
646 	struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
647 	const struct bpf_struct_ops *st_ops = st_map->st_ops;
648 	const struct btf_type *vt = st_ops->value_type;
649 	u64 usage;
650 
651 	usage = sizeof(*st_map) +
652 			vt->size - sizeof(struct bpf_struct_ops_value);
653 	usage += vt->size;
654 	usage += btf_type_vlen(vt) * sizeof(struct bpf_links *);
655 	usage += PAGE_SIZE;
656 	return usage;
657 }
658 
659 BTF_ID_LIST_SINGLE(bpf_struct_ops_map_btf_ids, struct, bpf_struct_ops_map)
660 const struct bpf_map_ops bpf_struct_ops_map_ops = {
661 	.map_alloc_check = bpf_struct_ops_map_alloc_check,
662 	.map_alloc = bpf_struct_ops_map_alloc,
663 	.map_free = bpf_struct_ops_map_free,
664 	.map_get_next_key = bpf_struct_ops_map_get_next_key,
665 	.map_lookup_elem = bpf_struct_ops_map_lookup_elem,
666 	.map_delete_elem = bpf_struct_ops_map_delete_elem,
667 	.map_update_elem = bpf_struct_ops_map_update_elem,
668 	.map_seq_show_elem = bpf_struct_ops_map_seq_show_elem,
669 	.map_mem_usage = bpf_struct_ops_map_mem_usage,
670 	.map_btf_id = &bpf_struct_ops_map_btf_ids[0],
671 };
672 
673 /* "const void *" because some subsystem is
674  * passing a const (e.g. const struct tcp_congestion_ops *)
675  */
676 bool bpf_struct_ops_get(const void *kdata)
677 {
678 	struct bpf_struct_ops_value *kvalue;
679 
680 	kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
681 
682 	return refcount_inc_not_zero(&kvalue->refcnt);
683 }
684 
685 static void bpf_struct_ops_put_rcu(struct rcu_head *head)
686 {
687 	struct bpf_struct_ops_map *st_map;
688 
689 	st_map = container_of(head, struct bpf_struct_ops_map, rcu);
690 	bpf_map_put(&st_map->map);
691 }
692 
693 void bpf_struct_ops_put(const void *kdata)
694 {
695 	struct bpf_struct_ops_value *kvalue;
696 
697 	kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
698 	if (refcount_dec_and_test(&kvalue->refcnt)) {
699 		struct bpf_struct_ops_map *st_map;
700 
701 		st_map = container_of(kvalue, struct bpf_struct_ops_map,
702 				      kvalue);
703 		/* The struct_ops's function may switch to another struct_ops.
704 		 *
705 		 * For example, bpf_tcp_cc_x->init() may switch to
706 		 * another tcp_cc_y by calling
707 		 * setsockopt(TCP_CONGESTION, "tcp_cc_y").
708 		 * During the switch,  bpf_struct_ops_put(tcp_cc_x) is called
709 		 * and its map->refcnt may reach 0 which then free its
710 		 * trampoline image while tcp_cc_x is still running.
711 		 *
712 		 * Thus, a rcu grace period is needed here.
713 		 */
714 		call_rcu(&st_map->rcu, bpf_struct_ops_put_rcu);
715 	}
716 }
717