xref: /openbmc/linux/net/ipv4/tcp_fastopen.c (revision d2168146)
1 #include <linux/err.h>
2 #include <linux/init.h>
3 #include <linux/kernel.h>
4 #include <linux/list.h>
5 #include <linux/tcp.h>
6 #include <linux/rcupdate.h>
7 #include <linux/rculist.h>
8 #include <net/inetpeer.h>
9 #include <net/tcp.h>
10 
11 int sysctl_tcp_fastopen __read_mostly = TFO_CLIENT_ENABLE;
12 
13 struct tcp_fastopen_context __rcu *tcp_fastopen_ctx;
14 
15 static DEFINE_SPINLOCK(tcp_fastopen_ctx_lock);
16 
17 void tcp_fastopen_init_key_once(bool publish)
18 {
19 	static u8 key[TCP_FASTOPEN_KEY_LENGTH];
20 
21 	/* tcp_fastopen_reset_cipher publishes the new context
22 	 * atomically, so we allow this race happening here.
23 	 *
24 	 * All call sites of tcp_fastopen_cookie_gen also check
25 	 * for a valid cookie, so this is an acceptable risk.
26 	 */
27 	if (net_get_random_once(key, sizeof(key)) && publish)
28 		tcp_fastopen_reset_cipher(key, sizeof(key));
29 }
30 
31 static void tcp_fastopen_ctx_free(struct rcu_head *head)
32 {
33 	struct tcp_fastopen_context *ctx =
34 	    container_of(head, struct tcp_fastopen_context, rcu);
35 	crypto_free_cipher(ctx->tfm);
36 	kfree(ctx);
37 }
38 
39 int tcp_fastopen_reset_cipher(void *key, unsigned int len)
40 {
41 	int err;
42 	struct tcp_fastopen_context *ctx, *octx;
43 
44 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
45 	if (!ctx)
46 		return -ENOMEM;
47 	ctx->tfm = crypto_alloc_cipher("aes", 0, 0);
48 
49 	if (IS_ERR(ctx->tfm)) {
50 		err = PTR_ERR(ctx->tfm);
51 error:		kfree(ctx);
52 		pr_err("TCP: TFO aes cipher alloc error: %d\n", err);
53 		return err;
54 	}
55 	err = crypto_cipher_setkey(ctx->tfm, key, len);
56 	if (err) {
57 		pr_err("TCP: TFO cipher key error: %d\n", err);
58 		crypto_free_cipher(ctx->tfm);
59 		goto error;
60 	}
61 	memcpy(ctx->key, key, len);
62 
63 	spin_lock(&tcp_fastopen_ctx_lock);
64 
65 	octx = rcu_dereference_protected(tcp_fastopen_ctx,
66 				lockdep_is_held(&tcp_fastopen_ctx_lock));
67 	rcu_assign_pointer(tcp_fastopen_ctx, ctx);
68 	spin_unlock(&tcp_fastopen_ctx_lock);
69 
70 	if (octx)
71 		call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
72 	return err;
73 }
74 
75 static bool __tcp_fastopen_cookie_gen(const void *path,
76 				      struct tcp_fastopen_cookie *foc)
77 {
78 	struct tcp_fastopen_context *ctx;
79 	bool ok = false;
80 
81 	tcp_fastopen_init_key_once(true);
82 
83 	rcu_read_lock();
84 	ctx = rcu_dereference(tcp_fastopen_ctx);
85 	if (ctx) {
86 		crypto_cipher_encrypt_one(ctx->tfm, foc->val, path);
87 		foc->len = TCP_FASTOPEN_COOKIE_SIZE;
88 		ok = true;
89 	}
90 	rcu_read_unlock();
91 	return ok;
92 }
93 
94 /* Generate the fastopen cookie by doing aes128 encryption on both
95  * the source and destination addresses. Pad 0s for IPv4 or IPv4-mapped-IPv6
96  * addresses. For the longer IPv6 addresses use CBC-MAC.
97  *
98  * XXX (TFO) - refactor when TCP_FASTOPEN_COOKIE_SIZE != AES_BLOCK_SIZE.
99  */
100 static bool tcp_fastopen_cookie_gen(struct request_sock *req,
101 				    struct sk_buff *syn,
102 				    struct tcp_fastopen_cookie *foc)
103 {
104 	if (req->rsk_ops->family == AF_INET) {
105 		const struct iphdr *iph = ip_hdr(syn);
106 
107 		__be32 path[4] = { iph->saddr, iph->daddr, 0, 0 };
108 		return __tcp_fastopen_cookie_gen(path, foc);
109 	}
110 
111 #if IS_ENABLED(CONFIG_IPV6)
112 	if (req->rsk_ops->family == AF_INET6) {
113 		const struct ipv6hdr *ip6h = ipv6_hdr(syn);
114 		struct tcp_fastopen_cookie tmp;
115 
116 		if (__tcp_fastopen_cookie_gen(&ip6h->saddr, &tmp)) {
117 			struct in6_addr *buf = (struct in6_addr *) tmp.val;
118 			int i = 4;
119 
120 			for (i = 0; i < 4; i++)
121 				buf->s6_addr32[i] ^= ip6h->daddr.s6_addr32[i];
122 			return __tcp_fastopen_cookie_gen(buf, foc);
123 		}
124 	}
125 #endif
126 	return false;
127 }
128 
129 static bool tcp_fastopen_create_child(struct sock *sk,
130 				      struct sk_buff *skb,
131 				      struct dst_entry *dst,
132 				      struct request_sock *req)
133 {
134 	struct tcp_sock *tp;
135 	struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
136 	struct sock *child;
137 
138 	req->num_retrans = 0;
139 	req->num_timeout = 0;
140 	req->sk = NULL;
141 
142 	child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
143 	if (child == NULL)
144 		return false;
145 
146 	spin_lock(&queue->fastopenq->lock);
147 	queue->fastopenq->qlen++;
148 	spin_unlock(&queue->fastopenq->lock);
149 
150 	/* Initialize the child socket. Have to fix some values to take
151 	 * into account the child is a Fast Open socket and is created
152 	 * only out of the bits carried in the SYN packet.
153 	 */
154 	tp = tcp_sk(child);
155 
156 	tp->fastopen_rsk = req;
157 	/* Do a hold on the listner sk so that if the listener is being
158 	 * closed, the child that has been accepted can live on and still
159 	 * access listen_lock.
160 	 */
161 	sock_hold(sk);
162 	tcp_rsk(req)->listener = sk;
163 
164 	/* RFC1323: The window in SYN & SYN/ACK segments is never
165 	 * scaled. So correct it appropriately.
166 	 */
167 	tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
168 
169 	/* Activate the retrans timer so that SYNACK can be retransmitted.
170 	 * The request socket is not added to the SYN table of the parent
171 	 * because it's been added to the accept queue directly.
172 	 */
173 	inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
174 				  TCP_TIMEOUT_INIT, TCP_RTO_MAX);
175 
176 	/* Add the child socket directly into the accept queue */
177 	inet_csk_reqsk_queue_add(sk, req, child);
178 
179 	/* Now finish processing the fastopen child socket. */
180 	inet_csk(child)->icsk_af_ops->rebuild_header(child);
181 	tcp_init_congestion_control(child);
182 	tcp_mtup_init(child);
183 	tcp_init_metrics(child);
184 	tcp_init_buffer_space(child);
185 
186 	/* Queue the data carried in the SYN packet. We need to first
187 	 * bump skb's refcnt because the caller will attempt to free it.
188 	 *
189 	 * XXX (TFO) - we honor a zero-payload TFO request for now,
190 	 * (any reason not to?) but no need to queue the skb since
191 	 * there is no data. How about SYN+FIN?
192 	 */
193 	if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1) {
194 		skb = skb_get(skb);
195 		skb_dst_drop(skb);
196 		__skb_pull(skb, tcp_hdr(skb)->doff * 4);
197 		skb_set_owner_r(skb, child);
198 		__skb_queue_tail(&child->sk_receive_queue, skb);
199 		tp->syn_data_acked = 1;
200 	}
201 	tcp_rsk(req)->rcv_nxt = tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
202 	sk->sk_data_ready(sk);
203 	bh_unlock_sock(child);
204 	sock_put(child);
205 	WARN_ON(req->sk == NULL);
206 	return true;
207 }
208 EXPORT_SYMBOL(tcp_fastopen_create_child);
209 
210 static bool tcp_fastopen_queue_check(struct sock *sk)
211 {
212 	struct fastopen_queue *fastopenq;
213 
214 	/* Make sure the listener has enabled fastopen, and we don't
215 	 * exceed the max # of pending TFO requests allowed before trying
216 	 * to validating the cookie in order to avoid burning CPU cycles
217 	 * unnecessarily.
218 	 *
219 	 * XXX (TFO) - The implication of checking the max_qlen before
220 	 * processing a cookie request is that clients can't differentiate
221 	 * between qlen overflow causing Fast Open to be disabled
222 	 * temporarily vs a server not supporting Fast Open at all.
223 	 */
224 	fastopenq = inet_csk(sk)->icsk_accept_queue.fastopenq;
225 	if (fastopenq == NULL || fastopenq->max_qlen == 0)
226 		return false;
227 
228 	if (fastopenq->qlen >= fastopenq->max_qlen) {
229 		struct request_sock *req1;
230 		spin_lock(&fastopenq->lock);
231 		req1 = fastopenq->rskq_rst_head;
232 		if ((req1 == NULL) || time_after(req1->expires, jiffies)) {
233 			spin_unlock(&fastopenq->lock);
234 			NET_INC_STATS_BH(sock_net(sk),
235 					 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
236 			return false;
237 		}
238 		fastopenq->rskq_rst_head = req1->dl_next;
239 		fastopenq->qlen--;
240 		spin_unlock(&fastopenq->lock);
241 		reqsk_free(req1);
242 	}
243 	return true;
244 }
245 
246 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
247  * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
248  * cookie request (foc->len == 0).
249  */
250 bool tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
251 		      struct request_sock *req,
252 		      struct tcp_fastopen_cookie *foc,
253 		      struct dst_entry *dst)
254 {
255 	struct tcp_fastopen_cookie valid_foc = { .len = -1 };
256 	bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
257 
258 	if (!((sysctl_tcp_fastopen & TFO_SERVER_ENABLE) &&
259 	      (syn_data || foc->len >= 0) &&
260 	      tcp_fastopen_queue_check(sk))) {
261 		foc->len = -1;
262 		return false;
263 	}
264 
265 	if (syn_data && (sysctl_tcp_fastopen & TFO_SERVER_COOKIE_NOT_REQD))
266 		goto fastopen;
267 
268 	if (tcp_fastopen_cookie_gen(req, skb, &valid_foc) &&
269 	    foc->len == TCP_FASTOPEN_COOKIE_SIZE &&
270 	    foc->len == valid_foc.len &&
271 	    !memcmp(foc->val, valid_foc.val, foc->len)) {
272 		/* Cookie is valid. Create a (full) child socket to accept
273 		 * the data in SYN before returning a SYN-ACK to ack the
274 		 * data. If we fail to create the socket, fall back and
275 		 * ack the ISN only but includes the same cookie.
276 		 *
277 		 * Note: Data-less SYN with valid cookie is allowed to send
278 		 * data in SYN_RECV state.
279 		 */
280 fastopen:
281 		if (tcp_fastopen_create_child(sk, skb, dst, req)) {
282 			foc->len = -1;
283 			NET_INC_STATS_BH(sock_net(sk),
284 					 LINUX_MIB_TCPFASTOPENPASSIVE);
285 			return true;
286 		}
287 	}
288 
289 	NET_INC_STATS_BH(sock_net(sk), foc->len ?
290 			 LINUX_MIB_TCPFASTOPENPASSIVEFAIL :
291 			 LINUX_MIB_TCPFASTOPENCOOKIEREQD);
292 	*foc = valid_foc;
293 	return false;
294 }
295 EXPORT_SYMBOL(tcp_try_fastopen);
296