xref: /openbmc/linux/net/ipv4/bpf_tcp_ca.c (revision bbaf1ff0)
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 extern 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 
24 static int bpf_tcp_ca_init(struct btf *btf)
25 {
26 	s32 type_id;
27 
28 	type_id = btf_find_by_name_kind(btf, "sock", BTF_KIND_STRUCT);
29 	if (type_id < 0)
30 		return -EINVAL;
31 	sock_id = type_id;
32 
33 	type_id = btf_find_by_name_kind(btf, "tcp_sock", BTF_KIND_STRUCT);
34 	if (type_id < 0)
35 		return -EINVAL;
36 	tcp_sock_id = type_id;
37 	tcp_sock_type = btf_type_by_id(btf, tcp_sock_id);
38 
39 	return 0;
40 }
41 
42 static bool is_unsupported(u32 member_offset)
43 {
44 	unsigned int i;
45 
46 	for (i = 0; i < ARRAY_SIZE(unsupported_ops); i++) {
47 		if (member_offset == unsupported_ops[i])
48 			return true;
49 	}
50 
51 	return false;
52 }
53 
54 extern struct btf *btf_vmlinux;
55 
56 static bool bpf_tcp_ca_is_valid_access(int off, int size,
57 				       enum bpf_access_type type,
58 				       const struct bpf_prog *prog,
59 				       struct bpf_insn_access_aux *info)
60 {
61 	if (!bpf_tracing_btf_ctx_access(off, size, type, prog, info))
62 		return false;
63 
64 	if (base_type(info->reg_type) == PTR_TO_BTF_ID &&
65 	    !bpf_type_has_unsafe_modifiers(info->reg_type) &&
66 	    info->btf_id == sock_id)
67 		/* promote it to tcp_sock */
68 		info->btf_id = tcp_sock_id;
69 
70 	return true;
71 }
72 
73 static int bpf_tcp_ca_btf_struct_access(struct bpf_verifier_log *log,
74 					const struct bpf_reg_state *reg,
75 					int off, int size)
76 {
77 	const struct btf_type *t;
78 	size_t end;
79 
80 	t = btf_type_by_id(reg->btf, reg->btf_id);
81 	if (t != tcp_sock_type) {
82 		bpf_log(log, "only read is supported\n");
83 		return -EACCES;
84 	}
85 
86 	switch (off) {
87 	case offsetof(struct sock, sk_pacing_rate):
88 		end = offsetofend(struct sock, sk_pacing_rate);
89 		break;
90 	case offsetof(struct sock, sk_pacing_status):
91 		end = offsetofend(struct sock, sk_pacing_status);
92 		break;
93 	case bpf_ctx_range(struct inet_connection_sock, icsk_ca_priv):
94 		end = offsetofend(struct inet_connection_sock, icsk_ca_priv);
95 		break;
96 	case offsetof(struct inet_connection_sock, icsk_ack.pending):
97 		end = offsetofend(struct inet_connection_sock,
98 				  icsk_ack.pending);
99 		break;
100 	case offsetof(struct tcp_sock, snd_cwnd):
101 		end = offsetofend(struct tcp_sock, snd_cwnd);
102 		break;
103 	case offsetof(struct tcp_sock, snd_cwnd_cnt):
104 		end = offsetofend(struct tcp_sock, snd_cwnd_cnt);
105 		break;
106 	case offsetof(struct tcp_sock, snd_ssthresh):
107 		end = offsetofend(struct tcp_sock, snd_ssthresh);
108 		break;
109 	case offsetof(struct tcp_sock, ecn_flags):
110 		end = offsetofend(struct tcp_sock, ecn_flags);
111 		break;
112 	case offsetof(struct tcp_sock, app_limited):
113 		end = offsetofend(struct tcp_sock, app_limited);
114 		break;
115 	default:
116 		bpf_log(log, "no write support to tcp_sock at off %d\n", off);
117 		return -EACCES;
118 	}
119 
120 	if (off + size > end) {
121 		bpf_log(log,
122 			"write access at off %d with size %d beyond the member of tcp_sock ended at %zu\n",
123 			off, size, end);
124 		return -EACCES;
125 	}
126 
127 	return 0;
128 }
129 
130 BPF_CALL_2(bpf_tcp_send_ack, struct tcp_sock *, tp, u32, rcv_nxt)
131 {
132 	/* bpf_tcp_ca prog cannot have NULL tp */
133 	__tcp_send_ack((struct sock *)tp, rcv_nxt);
134 	return 0;
135 }
136 
137 static const struct bpf_func_proto bpf_tcp_send_ack_proto = {
138 	.func		= bpf_tcp_send_ack,
139 	.gpl_only	= false,
140 	/* In case we want to report error later */
141 	.ret_type	= RET_INTEGER,
142 	.arg1_type	= ARG_PTR_TO_BTF_ID,
143 	.arg1_btf_id	= &tcp_sock_id,
144 	.arg2_type	= ARG_ANYTHING,
145 };
146 
147 static u32 prog_ops_moff(const struct bpf_prog *prog)
148 {
149 	const struct btf_member *m;
150 	const struct btf_type *t;
151 	u32 midx;
152 
153 	midx = prog->expected_attach_type;
154 	t = bpf_tcp_congestion_ops.type;
155 	m = &btf_type_member(t)[midx];
156 
157 	return __btf_member_bit_offset(t, m) / 8;
158 }
159 
160 static const struct bpf_func_proto *
161 bpf_tcp_ca_get_func_proto(enum bpf_func_id func_id,
162 			  const struct bpf_prog *prog)
163 {
164 	switch (func_id) {
165 	case BPF_FUNC_tcp_send_ack:
166 		return &bpf_tcp_send_ack_proto;
167 	case BPF_FUNC_sk_storage_get:
168 		return &bpf_sk_storage_get_proto;
169 	case BPF_FUNC_sk_storage_delete:
170 		return &bpf_sk_storage_delete_proto;
171 	case BPF_FUNC_setsockopt:
172 		/* Does not allow release() to call setsockopt.
173 		 * release() is called when the current bpf-tcp-cc
174 		 * is retiring.  It is not allowed to call
175 		 * setsockopt() to make further changes which
176 		 * may potentially allocate new resources.
177 		 */
178 		if (prog_ops_moff(prog) !=
179 		    offsetof(struct tcp_congestion_ops, release))
180 			return &bpf_sk_setsockopt_proto;
181 		return NULL;
182 	case BPF_FUNC_getsockopt:
183 		/* Since get/setsockopt is usually expected to
184 		 * be available together, disable getsockopt for
185 		 * release also to avoid usage surprise.
186 		 * The bpf-tcp-cc already has a more powerful way
187 		 * to read tcp_sock from the PTR_TO_BTF_ID.
188 		 */
189 		if (prog_ops_moff(prog) !=
190 		    offsetof(struct tcp_congestion_ops, release))
191 			return &bpf_sk_getsockopt_proto;
192 		return NULL;
193 	case BPF_FUNC_ktime_get_coarse_ns:
194 		return &bpf_ktime_get_coarse_ns_proto;
195 	default:
196 		return bpf_base_func_proto(func_id);
197 	}
198 }
199 
200 BTF_SET8_START(bpf_tcp_ca_check_kfunc_ids)
201 BTF_ID_FLAGS(func, tcp_reno_ssthresh)
202 BTF_ID_FLAGS(func, tcp_reno_cong_avoid)
203 BTF_ID_FLAGS(func, tcp_reno_undo_cwnd)
204 BTF_ID_FLAGS(func, tcp_slow_start)
205 BTF_ID_FLAGS(func, tcp_cong_avoid_ai)
206 BTF_SET8_END(bpf_tcp_ca_check_kfunc_ids)
207 
208 static const struct btf_kfunc_id_set bpf_tcp_ca_kfunc_set = {
209 	.owner = THIS_MODULE,
210 	.set   = &bpf_tcp_ca_check_kfunc_ids,
211 };
212 
213 static const struct bpf_verifier_ops bpf_tcp_ca_verifier_ops = {
214 	.get_func_proto		= bpf_tcp_ca_get_func_proto,
215 	.is_valid_access	= bpf_tcp_ca_is_valid_access,
216 	.btf_struct_access	= bpf_tcp_ca_btf_struct_access,
217 };
218 
219 static int bpf_tcp_ca_init_member(const struct btf_type *t,
220 				  const struct btf_member *member,
221 				  void *kdata, const void *udata)
222 {
223 	const struct tcp_congestion_ops *utcp_ca;
224 	struct tcp_congestion_ops *tcp_ca;
225 	u32 moff;
226 
227 	utcp_ca = (const struct tcp_congestion_ops *)udata;
228 	tcp_ca = (struct tcp_congestion_ops *)kdata;
229 
230 	moff = __btf_member_bit_offset(t, member) / 8;
231 	switch (moff) {
232 	case offsetof(struct tcp_congestion_ops, flags):
233 		if (utcp_ca->flags & ~TCP_CONG_MASK)
234 			return -EINVAL;
235 		tcp_ca->flags = utcp_ca->flags;
236 		return 1;
237 	case offsetof(struct tcp_congestion_ops, name):
238 		if (bpf_obj_name_cpy(tcp_ca->name, utcp_ca->name,
239 				     sizeof(tcp_ca->name)) <= 0)
240 			return -EINVAL;
241 		return 1;
242 	}
243 
244 	return 0;
245 }
246 
247 static int bpf_tcp_ca_check_member(const struct btf_type *t,
248 				   const struct btf_member *member,
249 				   const struct bpf_prog *prog)
250 {
251 	if (is_unsupported(__btf_member_bit_offset(t, member) / 8))
252 		return -ENOTSUPP;
253 	return 0;
254 }
255 
256 static int bpf_tcp_ca_reg(void *kdata)
257 {
258 	return tcp_register_congestion_control(kdata);
259 }
260 
261 static void bpf_tcp_ca_unreg(void *kdata)
262 {
263 	tcp_unregister_congestion_control(kdata);
264 }
265 
266 static int bpf_tcp_ca_update(void *kdata, void *old_kdata)
267 {
268 	return tcp_update_congestion_control(kdata, old_kdata);
269 }
270 
271 static int bpf_tcp_ca_validate(void *kdata)
272 {
273 	return tcp_validate_congestion_control(kdata);
274 }
275 
276 struct bpf_struct_ops bpf_tcp_congestion_ops = {
277 	.verifier_ops = &bpf_tcp_ca_verifier_ops,
278 	.reg = bpf_tcp_ca_reg,
279 	.unreg = bpf_tcp_ca_unreg,
280 	.update = bpf_tcp_ca_update,
281 	.check_member = bpf_tcp_ca_check_member,
282 	.init_member = bpf_tcp_ca_init_member,
283 	.init = bpf_tcp_ca_init,
284 	.validate = bpf_tcp_ca_validate,
285 	.name = "tcp_congestion_ops",
286 };
287 
288 static int __init bpf_tcp_ca_kfunc_init(void)
289 {
290 	return register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &bpf_tcp_ca_kfunc_set);
291 }
292 late_initcall(bpf_tcp_ca_kfunc_init);
293