xref: /openbmc/linux/net/ipv4/bpf_tcp_ca.c (revision ecefa105)
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, enum bpf_access_type atype,
76 					u32 *next_btf_id, enum bpf_type_flag *flag)
77 {
78 	const struct btf_type *t;
79 	size_t end;
80 
81 	if (atype == BPF_READ)
82 		return btf_struct_access(log, reg, off, size, atype, next_btf_id, flag);
83 
84 	t = btf_type_by_id(reg->btf, reg->btf_id);
85 	if (t != tcp_sock_type) {
86 		bpf_log(log, "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 	default:
117 		bpf_log(log, "no write support to tcp_sock at off %d\n", off);
118 		return -EACCES;
119 	}
120 
121 	if (off + size > end) {
122 		bpf_log(log,
123 			"write access at off %d with size %d beyond the member of tcp_sock ended at %zu\n",
124 			off, size, end);
125 		return -EACCES;
126 	}
127 
128 	return 0;
129 }
130 
131 BPF_CALL_2(bpf_tcp_send_ack, struct tcp_sock *, tp, u32, rcv_nxt)
132 {
133 	/* bpf_tcp_ca prog cannot have NULL tp */
134 	__tcp_send_ack((struct sock *)tp, rcv_nxt);
135 	return 0;
136 }
137 
138 static const struct bpf_func_proto bpf_tcp_send_ack_proto = {
139 	.func		= bpf_tcp_send_ack,
140 	.gpl_only	= false,
141 	/* In case we want to report error later */
142 	.ret_type	= RET_INTEGER,
143 	.arg1_type	= ARG_PTR_TO_BTF_ID,
144 	.arg1_btf_id	= &tcp_sock_id,
145 	.arg2_type	= ARG_ANYTHING,
146 };
147 
148 static u32 prog_ops_moff(const struct bpf_prog *prog)
149 {
150 	const struct btf_member *m;
151 	const struct btf_type *t;
152 	u32 midx;
153 
154 	midx = prog->expected_attach_type;
155 	t = bpf_tcp_congestion_ops.type;
156 	m = &btf_type_member(t)[midx];
157 
158 	return __btf_member_bit_offset(t, m) / 8;
159 }
160 
161 static const struct bpf_func_proto *
162 bpf_tcp_ca_get_func_proto(enum bpf_func_id func_id,
163 			  const struct bpf_prog *prog)
164 {
165 	switch (func_id) {
166 	case BPF_FUNC_tcp_send_ack:
167 		return &bpf_tcp_send_ack_proto;
168 	case BPF_FUNC_sk_storage_get:
169 		return &bpf_sk_storage_get_proto;
170 	case BPF_FUNC_sk_storage_delete:
171 		return &bpf_sk_storage_delete_proto;
172 	case BPF_FUNC_setsockopt:
173 		/* Does not allow release() to call setsockopt.
174 		 * release() is called when the current bpf-tcp-cc
175 		 * is retiring.  It is not allowed to call
176 		 * setsockopt() to make further changes which
177 		 * may potentially allocate new resources.
178 		 */
179 		if (prog_ops_moff(prog) !=
180 		    offsetof(struct tcp_congestion_ops, release))
181 			return &bpf_sk_setsockopt_proto;
182 		return NULL;
183 	case BPF_FUNC_getsockopt:
184 		/* Since get/setsockopt is usually expected to
185 		 * be available together, disable getsockopt for
186 		 * release also to avoid usage surprise.
187 		 * The bpf-tcp-cc already has a more powerful way
188 		 * to read tcp_sock from the PTR_TO_BTF_ID.
189 		 */
190 		if (prog_ops_moff(prog) !=
191 		    offsetof(struct tcp_congestion_ops, release))
192 			return &bpf_sk_getsockopt_proto;
193 		return NULL;
194 	case BPF_FUNC_ktime_get_coarse_ns:
195 		return &bpf_ktime_get_coarse_ns_proto;
196 	default:
197 		return bpf_base_func_proto(func_id);
198 	}
199 }
200 
201 BTF_SET8_START(bpf_tcp_ca_check_kfunc_ids)
202 BTF_ID_FLAGS(func, tcp_reno_ssthresh)
203 BTF_ID_FLAGS(func, tcp_reno_cong_avoid)
204 BTF_ID_FLAGS(func, tcp_reno_undo_cwnd)
205 BTF_ID_FLAGS(func, tcp_slow_start)
206 BTF_ID_FLAGS(func, tcp_cong_avoid_ai)
207 BTF_SET8_END(bpf_tcp_ca_check_kfunc_ids)
208 
209 static const struct btf_kfunc_id_set bpf_tcp_ca_kfunc_set = {
210 	.owner = THIS_MODULE,
211 	.set   = &bpf_tcp_ca_check_kfunc_ids,
212 };
213 
214 static const struct bpf_verifier_ops bpf_tcp_ca_verifier_ops = {
215 	.get_func_proto		= bpf_tcp_ca_get_func_proto,
216 	.is_valid_access	= bpf_tcp_ca_is_valid_access,
217 	.btf_struct_access	= bpf_tcp_ca_btf_struct_access,
218 };
219 
220 static int bpf_tcp_ca_init_member(const struct btf_type *t,
221 				  const struct btf_member *member,
222 				  void *kdata, const void *udata)
223 {
224 	const struct tcp_congestion_ops *utcp_ca;
225 	struct tcp_congestion_ops *tcp_ca;
226 	u32 moff;
227 
228 	utcp_ca = (const struct tcp_congestion_ops *)udata;
229 	tcp_ca = (struct tcp_congestion_ops *)kdata;
230 
231 	moff = __btf_member_bit_offset(t, member) / 8;
232 	switch (moff) {
233 	case offsetof(struct tcp_congestion_ops, flags):
234 		if (utcp_ca->flags & ~TCP_CONG_MASK)
235 			return -EINVAL;
236 		tcp_ca->flags = utcp_ca->flags;
237 		return 1;
238 	case offsetof(struct tcp_congestion_ops, name):
239 		if (bpf_obj_name_cpy(tcp_ca->name, utcp_ca->name,
240 				     sizeof(tcp_ca->name)) <= 0)
241 			return -EINVAL;
242 		if (tcp_ca_find(utcp_ca->name))
243 			return -EEXIST;
244 		return 1;
245 	}
246 
247 	return 0;
248 }
249 
250 static int bpf_tcp_ca_check_member(const struct btf_type *t,
251 				   const struct btf_member *member,
252 				   const struct bpf_prog *prog)
253 {
254 	if (is_unsupported(__btf_member_bit_offset(t, member) / 8))
255 		return -ENOTSUPP;
256 	return 0;
257 }
258 
259 static int bpf_tcp_ca_reg(void *kdata)
260 {
261 	return tcp_register_congestion_control(kdata);
262 }
263 
264 static void bpf_tcp_ca_unreg(void *kdata)
265 {
266 	tcp_unregister_congestion_control(kdata);
267 }
268 
269 struct bpf_struct_ops bpf_tcp_congestion_ops = {
270 	.verifier_ops = &bpf_tcp_ca_verifier_ops,
271 	.reg = bpf_tcp_ca_reg,
272 	.unreg = bpf_tcp_ca_unreg,
273 	.check_member = bpf_tcp_ca_check_member,
274 	.init_member = bpf_tcp_ca_init_member,
275 	.init = bpf_tcp_ca_init,
276 	.name = "tcp_congestion_ops",
277 };
278 
279 static int __init bpf_tcp_ca_kfunc_init(void)
280 {
281 	return register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &bpf_tcp_ca_kfunc_set);
282 }
283 late_initcall(bpf_tcp_ca_kfunc_init);
284