1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* Copyright (c) 2019 Facebook */ |
3 | |
4 | #include <linux/init.h> |
5 | #include <linux/types.h> |
6 | #include <linux/bpf_verifier.h> |
7 | #include <linux/bpf.h> |
8 | #include <linux/btf.h> |
9 | #include <linux/btf_ids.h> |
10 | #include <linux/filter.h> |
11 | #include <net/tcp.h> |
12 | #include <net/bpf_sk_storage.h> |
13 | |
14 | /* "extern" is to avoid sparse warning. It is only used in bpf_struct_ops.c. */ |
15 | static struct bpf_struct_ops bpf_tcp_congestion_ops; |
16 | |
17 | static u32 unsupported_ops[] = { |
18 | offsetof(struct tcp_congestion_ops, get_info), |
19 | }; |
20 | |
21 | static const struct btf_type *tcp_sock_type; |
22 | static u32 tcp_sock_id, sock_id; |
23 | static const struct btf_type *tcp_congestion_ops_type; |
24 | |
25 | static int bpf_tcp_ca_init(struct btf *btf) |
26 | { |
27 | s32 type_id; |
28 | |
29 | type_id = btf_find_by_name_kind(btf, name: "sock" , kind: BTF_KIND_STRUCT); |
30 | if (type_id < 0) |
31 | return -EINVAL; |
32 | sock_id = type_id; |
33 | |
34 | type_id = btf_find_by_name_kind(btf, name: "tcp_sock" , kind: BTF_KIND_STRUCT); |
35 | if (type_id < 0) |
36 | return -EINVAL; |
37 | tcp_sock_id = type_id; |
38 | tcp_sock_type = btf_type_by_id(btf, type_id: tcp_sock_id); |
39 | |
40 | type_id = btf_find_by_name_kind(btf, name: "tcp_congestion_ops" , kind: BTF_KIND_STRUCT); |
41 | if (type_id < 0) |
42 | return -EINVAL; |
43 | tcp_congestion_ops_type = btf_type_by_id(btf, type_id); |
44 | |
45 | return 0; |
46 | } |
47 | |
48 | static bool is_unsupported(u32 member_offset) |
49 | { |
50 | unsigned int i; |
51 | |
52 | for (i = 0; i < ARRAY_SIZE(unsupported_ops); i++) { |
53 | if (member_offset == unsupported_ops[i]) |
54 | return true; |
55 | } |
56 | |
57 | return false; |
58 | } |
59 | |
60 | static bool bpf_tcp_ca_is_valid_access(int off, int size, |
61 | enum bpf_access_type type, |
62 | const struct bpf_prog *prog, |
63 | struct bpf_insn_access_aux *info) |
64 | { |
65 | if (!bpf_tracing_btf_ctx_access(off, size, type, prog, info)) |
66 | return false; |
67 | |
68 | if (base_type(type: info->reg_type) == PTR_TO_BTF_ID && |
69 | !bpf_type_has_unsafe_modifiers(type: info->reg_type) && |
70 | info->btf_id == sock_id) |
71 | /* promote it to tcp_sock */ |
72 | info->btf_id = tcp_sock_id; |
73 | |
74 | return true; |
75 | } |
76 | |
77 | static int bpf_tcp_ca_btf_struct_access(struct bpf_verifier_log *log, |
78 | const struct bpf_reg_state *reg, |
79 | int off, int size) |
80 | { |
81 | const struct btf_type *t; |
82 | size_t end; |
83 | |
84 | t = btf_type_by_id(btf: reg->btf, type_id: reg->btf_id); |
85 | if (t != tcp_sock_type) { |
86 | bpf_log(log, fmt: "only read is supported\n" ); |
87 | return -EACCES; |
88 | } |
89 | |
90 | switch (off) { |
91 | case offsetof(struct sock, sk_pacing_rate): |
92 | end = offsetofend(struct sock, sk_pacing_rate); |
93 | break; |
94 | case offsetof(struct sock, sk_pacing_status): |
95 | end = offsetofend(struct sock, sk_pacing_status); |
96 | break; |
97 | case bpf_ctx_range(struct inet_connection_sock, icsk_ca_priv): |
98 | end = offsetofend(struct inet_connection_sock, icsk_ca_priv); |
99 | break; |
100 | case offsetof(struct inet_connection_sock, icsk_ack.pending): |
101 | end = offsetofend(struct inet_connection_sock, |
102 | icsk_ack.pending); |
103 | break; |
104 | case offsetof(struct tcp_sock, snd_cwnd): |
105 | end = offsetofend(struct tcp_sock, snd_cwnd); |
106 | break; |
107 | case offsetof(struct tcp_sock, snd_cwnd_cnt): |
108 | end = offsetofend(struct tcp_sock, snd_cwnd_cnt); |
109 | break; |
110 | case offsetof(struct tcp_sock, snd_ssthresh): |
111 | end = offsetofend(struct tcp_sock, snd_ssthresh); |
112 | break; |
113 | case offsetof(struct tcp_sock, ecn_flags): |
114 | end = offsetofend(struct tcp_sock, ecn_flags); |
115 | break; |
116 | case offsetof(struct tcp_sock, app_limited): |
117 | end = offsetofend(struct tcp_sock, app_limited); |
118 | break; |
119 | default: |
120 | bpf_log(log, fmt: "no write support to tcp_sock at off %d\n" , off); |
121 | return -EACCES; |
122 | } |
123 | |
124 | if (off + size > end) { |
125 | bpf_log(log, |
126 | fmt: "write access at off %d with size %d beyond the member of tcp_sock ended at %zu\n" , |
127 | off, size, end); |
128 | return -EACCES; |
129 | } |
130 | |
131 | return 0; |
132 | } |
133 | |
134 | BPF_CALL_2(bpf_tcp_send_ack, struct tcp_sock *, tp, u32, rcv_nxt) |
135 | { |
136 | /* bpf_tcp_ca prog cannot have NULL tp */ |
137 | __tcp_send_ack(sk: (struct sock *)tp, rcv_nxt); |
138 | return 0; |
139 | } |
140 | |
141 | static const struct bpf_func_proto bpf_tcp_send_ack_proto = { |
142 | .func = bpf_tcp_send_ack, |
143 | .gpl_only = false, |
144 | /* In case we want to report error later */ |
145 | .ret_type = RET_INTEGER, |
146 | .arg1_type = ARG_PTR_TO_BTF_ID, |
147 | .arg1_btf_id = &tcp_sock_id, |
148 | .arg2_type = ARG_ANYTHING, |
149 | }; |
150 | |
151 | static u32 prog_ops_moff(const struct bpf_prog *prog) |
152 | { |
153 | const struct btf_member *m; |
154 | const struct btf_type *t; |
155 | u32 midx; |
156 | |
157 | midx = prog->expected_attach_type; |
158 | t = tcp_congestion_ops_type; |
159 | m = &btf_type_member(t)[midx]; |
160 | |
161 | return __btf_member_bit_offset(struct_type: t, member: m) / 8; |
162 | } |
163 | |
164 | static const struct bpf_func_proto * |
165 | bpf_tcp_ca_get_func_proto(enum bpf_func_id func_id, |
166 | const struct bpf_prog *prog) |
167 | { |
168 | switch (func_id) { |
169 | case BPF_FUNC_tcp_send_ack: |
170 | return &bpf_tcp_send_ack_proto; |
171 | case BPF_FUNC_sk_storage_get: |
172 | return &bpf_sk_storage_get_proto; |
173 | case BPF_FUNC_sk_storage_delete: |
174 | return &bpf_sk_storage_delete_proto; |
175 | case BPF_FUNC_setsockopt: |
176 | /* Does not allow release() to call setsockopt. |
177 | * release() is called when the current bpf-tcp-cc |
178 | * is retiring. It is not allowed to call |
179 | * setsockopt() to make further changes which |
180 | * may potentially allocate new resources. |
181 | */ |
182 | if (prog_ops_moff(prog) != |
183 | offsetof(struct tcp_congestion_ops, release)) |
184 | return &bpf_sk_setsockopt_proto; |
185 | return NULL; |
186 | case BPF_FUNC_getsockopt: |
187 | /* Since get/setsockopt is usually expected to |
188 | * be available together, disable getsockopt for |
189 | * release also to avoid usage surprise. |
190 | * The bpf-tcp-cc already has a more powerful way |
191 | * to read tcp_sock from the PTR_TO_BTF_ID. |
192 | */ |
193 | if (prog_ops_moff(prog) != |
194 | offsetof(struct tcp_congestion_ops, release)) |
195 | return &bpf_sk_getsockopt_proto; |
196 | return NULL; |
197 | case BPF_FUNC_ktime_get_coarse_ns: |
198 | return &bpf_ktime_get_coarse_ns_proto; |
199 | default: |
200 | return bpf_base_func_proto(func_id, prog); |
201 | } |
202 | } |
203 | |
204 | BTF_KFUNCS_START(bpf_tcp_ca_check_kfunc_ids) |
205 | BTF_ID_FLAGS(func, tcp_reno_ssthresh) |
206 | BTF_ID_FLAGS(func, tcp_reno_cong_avoid) |
207 | BTF_ID_FLAGS(func, tcp_reno_undo_cwnd) |
208 | BTF_ID_FLAGS(func, tcp_slow_start) |
209 | BTF_ID_FLAGS(func, tcp_cong_avoid_ai) |
210 | BTF_KFUNCS_END(bpf_tcp_ca_check_kfunc_ids) |
211 | |
212 | static const struct btf_kfunc_id_set bpf_tcp_ca_kfunc_set = { |
213 | .owner = THIS_MODULE, |
214 | .set = &bpf_tcp_ca_check_kfunc_ids, |
215 | }; |
216 | |
217 | static const struct bpf_verifier_ops bpf_tcp_ca_verifier_ops = { |
218 | .get_func_proto = bpf_tcp_ca_get_func_proto, |
219 | .is_valid_access = bpf_tcp_ca_is_valid_access, |
220 | .btf_struct_access = bpf_tcp_ca_btf_struct_access, |
221 | }; |
222 | |
223 | static int bpf_tcp_ca_init_member(const struct btf_type *t, |
224 | const struct btf_member *member, |
225 | void *kdata, const void *udata) |
226 | { |
227 | const struct tcp_congestion_ops *utcp_ca; |
228 | struct tcp_congestion_ops *tcp_ca; |
229 | u32 moff; |
230 | |
231 | utcp_ca = (const struct tcp_congestion_ops *)udata; |
232 | tcp_ca = (struct tcp_congestion_ops *)kdata; |
233 | |
234 | moff = __btf_member_bit_offset(struct_type: t, member) / 8; |
235 | switch (moff) { |
236 | case offsetof(struct tcp_congestion_ops, flags): |
237 | if (utcp_ca->flags & ~TCP_CONG_MASK) |
238 | return -EINVAL; |
239 | tcp_ca->flags = utcp_ca->flags; |
240 | return 1; |
241 | case offsetof(struct tcp_congestion_ops, name): |
242 | if (bpf_obj_name_cpy(dst: tcp_ca->name, src: utcp_ca->name, |
243 | size: sizeof(tcp_ca->name)) <= 0) |
244 | return -EINVAL; |
245 | return 1; |
246 | } |
247 | |
248 | return 0; |
249 | } |
250 | |
251 | static int bpf_tcp_ca_check_member(const struct btf_type *t, |
252 | const struct btf_member *member, |
253 | const struct bpf_prog *prog) |
254 | { |
255 | if (is_unsupported(member_offset: __btf_member_bit_offset(struct_type: t, member) / 8)) |
256 | return -ENOTSUPP; |
257 | return 0; |
258 | } |
259 | |
260 | static int bpf_tcp_ca_reg(void *kdata) |
261 | { |
262 | return tcp_register_congestion_control(type: kdata); |
263 | } |
264 | |
265 | static void bpf_tcp_ca_unreg(void *kdata) |
266 | { |
267 | tcp_unregister_congestion_control(type: kdata); |
268 | } |
269 | |
270 | static int bpf_tcp_ca_update(void *kdata, void *old_kdata) |
271 | { |
272 | return tcp_update_congestion_control(type: kdata, old_type: old_kdata); |
273 | } |
274 | |
275 | static int bpf_tcp_ca_validate(void *kdata) |
276 | { |
277 | return tcp_validate_congestion_control(ca: kdata); |
278 | } |
279 | |
280 | static u32 bpf_tcp_ca_ssthresh(struct sock *sk) |
281 | { |
282 | return 0; |
283 | } |
284 | |
285 | static void bpf_tcp_ca_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
286 | { |
287 | } |
288 | |
289 | static void bpf_tcp_ca_set_state(struct sock *sk, u8 new_state) |
290 | { |
291 | } |
292 | |
293 | static void bpf_tcp_ca_cwnd_event(struct sock *sk, enum tcp_ca_event ev) |
294 | { |
295 | } |
296 | |
297 | static void bpf_tcp_ca_in_ack_event(struct sock *sk, u32 flags) |
298 | { |
299 | } |
300 | |
301 | static void bpf_tcp_ca_pkts_acked(struct sock *sk, const struct ack_sample *sample) |
302 | { |
303 | } |
304 | |
305 | static u32 bpf_tcp_ca_min_tso_segs(struct sock *sk) |
306 | { |
307 | return 0; |
308 | } |
309 | |
310 | static void bpf_tcp_ca_cong_control(struct sock *sk, const struct rate_sample *rs) |
311 | { |
312 | } |
313 | |
314 | static u32 bpf_tcp_ca_undo_cwnd(struct sock *sk) |
315 | { |
316 | return 0; |
317 | } |
318 | |
319 | static u32 bpf_tcp_ca_sndbuf_expand(struct sock *sk) |
320 | { |
321 | return 0; |
322 | } |
323 | |
324 | static void __bpf_tcp_ca_init(struct sock *sk) |
325 | { |
326 | } |
327 | |
328 | static void __bpf_tcp_ca_release(struct sock *sk) |
329 | { |
330 | } |
331 | |
332 | static struct tcp_congestion_ops __bpf_ops_tcp_congestion_ops = { |
333 | .ssthresh = bpf_tcp_ca_ssthresh, |
334 | .cong_avoid = bpf_tcp_ca_cong_avoid, |
335 | .set_state = bpf_tcp_ca_set_state, |
336 | .cwnd_event = bpf_tcp_ca_cwnd_event, |
337 | .in_ack_event = bpf_tcp_ca_in_ack_event, |
338 | .pkts_acked = bpf_tcp_ca_pkts_acked, |
339 | .min_tso_segs = bpf_tcp_ca_min_tso_segs, |
340 | .cong_control = bpf_tcp_ca_cong_control, |
341 | .undo_cwnd = bpf_tcp_ca_undo_cwnd, |
342 | .sndbuf_expand = bpf_tcp_ca_sndbuf_expand, |
343 | |
344 | .init = __bpf_tcp_ca_init, |
345 | .release = __bpf_tcp_ca_release, |
346 | }; |
347 | |
348 | static struct bpf_struct_ops bpf_tcp_congestion_ops = { |
349 | .verifier_ops = &bpf_tcp_ca_verifier_ops, |
350 | .reg = bpf_tcp_ca_reg, |
351 | .unreg = bpf_tcp_ca_unreg, |
352 | .update = bpf_tcp_ca_update, |
353 | .check_member = bpf_tcp_ca_check_member, |
354 | .init_member = bpf_tcp_ca_init_member, |
355 | .init = bpf_tcp_ca_init, |
356 | .validate = bpf_tcp_ca_validate, |
357 | .name = "tcp_congestion_ops" , |
358 | .cfi_stubs = &__bpf_ops_tcp_congestion_ops, |
359 | .owner = THIS_MODULE, |
360 | }; |
361 | |
362 | static int __init bpf_tcp_ca_kfunc_init(void) |
363 | { |
364 | int ret; |
365 | |
366 | ret = register_btf_kfunc_id_set(prog_type: BPF_PROG_TYPE_STRUCT_OPS, s: &bpf_tcp_ca_kfunc_set); |
367 | ret = ret ?: register_bpf_struct_ops(&bpf_tcp_congestion_ops, tcp_congestion_ops); |
368 | |
369 | return ret; |
370 | } |
371 | late_initcall(bpf_tcp_ca_kfunc_init); |
372 | |