xref: /openbmc/linux/net/ipv4/tcp_fastopen.c (revision 827634ad)
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;
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 	u32 end_seq;
138 
139 	req->num_retrans = 0;
140 	req->num_timeout = 0;
141 	req->sk = NULL;
142 
143 	child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
144 	if (!child)
145 		return false;
146 
147 	spin_lock(&queue->fastopenq->lock);
148 	queue->fastopenq->qlen++;
149 	spin_unlock(&queue->fastopenq->lock);
150 
151 	/* Initialize the child socket. Have to fix some values to take
152 	 * into account the child is a Fast Open socket and is created
153 	 * only out of the bits carried in the SYN packet.
154 	 */
155 	tp = tcp_sk(child);
156 
157 	tp->fastopen_rsk = req;
158 	tcp_rsk(req)->tfo_listener = true;
159 
160 	/* RFC1323: The window in SYN & SYN/ACK segments is never
161 	 * scaled. So correct it appropriately.
162 	 */
163 	tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
164 
165 	/* Activate the retrans timer so that SYNACK can be retransmitted.
166 	 * The request socket is not added to the SYN table of the parent
167 	 * because it's been added to the accept queue directly.
168 	 */
169 	inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
170 				  TCP_TIMEOUT_INIT, TCP_RTO_MAX);
171 
172 	atomic_set(&req->rsk_refcnt, 1);
173 	/* Add the child socket directly into the accept queue */
174 	inet_csk_reqsk_queue_add(sk, req, child);
175 
176 	/* Now finish processing the fastopen child socket. */
177 	inet_csk(child)->icsk_af_ops->rebuild_header(child);
178 	tcp_init_congestion_control(child);
179 	tcp_mtup_init(child);
180 	tcp_init_metrics(child);
181 	tcp_init_buffer_space(child);
182 
183 	/* Queue the data carried in the SYN packet. We need to first
184 	 * bump skb's refcnt because the caller will attempt to free it.
185 	 * Note that IPv6 might also have used skb_get() trick
186 	 * in tcp_v6_conn_request() to keep this SYN around (treq->pktopts)
187 	 * So we need to eventually get a clone of the packet,
188 	 * before inserting it in sk_receive_queue.
189 	 *
190 	 * XXX (TFO) - we honor a zero-payload TFO request for now,
191 	 * (any reason not to?) but no need to queue the skb since
192 	 * there is no data. How about SYN+FIN?
193 	 */
194 	end_seq = TCP_SKB_CB(skb)->end_seq;
195 	if (end_seq != TCP_SKB_CB(skb)->seq + 1) {
196 		struct sk_buff *skb2;
197 
198 		if (unlikely(skb_shared(skb)))
199 			skb2 = skb_clone(skb, GFP_ATOMIC);
200 		else
201 			skb2 = skb_get(skb);
202 
203 		if (likely(skb2)) {
204 			skb_dst_drop(skb2);
205 			__skb_pull(skb2, tcp_hdrlen(skb));
206 			skb_set_owner_r(skb2, child);
207 			__skb_queue_tail(&child->sk_receive_queue, skb2);
208 			tp->syn_data_acked = 1;
209 		} else {
210 			end_seq = TCP_SKB_CB(skb)->seq + 1;
211 		}
212 	}
213 	tcp_rsk(req)->rcv_nxt = tp->rcv_nxt = end_seq;
214 	sk->sk_data_ready(sk);
215 	bh_unlock_sock(child);
216 	sock_put(child);
217 	WARN_ON(!req->sk);
218 	return true;
219 }
220 
221 static bool tcp_fastopen_queue_check(struct sock *sk)
222 {
223 	struct fastopen_queue *fastopenq;
224 
225 	/* Make sure the listener has enabled fastopen, and we don't
226 	 * exceed the max # of pending TFO requests allowed before trying
227 	 * to validating the cookie in order to avoid burning CPU cycles
228 	 * unnecessarily.
229 	 *
230 	 * XXX (TFO) - The implication of checking the max_qlen before
231 	 * processing a cookie request is that clients can't differentiate
232 	 * between qlen overflow causing Fast Open to be disabled
233 	 * temporarily vs a server not supporting Fast Open at all.
234 	 */
235 	fastopenq = inet_csk(sk)->icsk_accept_queue.fastopenq;
236 	if (!fastopenq || fastopenq->max_qlen == 0)
237 		return false;
238 
239 	if (fastopenq->qlen >= fastopenq->max_qlen) {
240 		struct request_sock *req1;
241 		spin_lock(&fastopenq->lock);
242 		req1 = fastopenq->rskq_rst_head;
243 		if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
244 			spin_unlock(&fastopenq->lock);
245 			NET_INC_STATS_BH(sock_net(sk),
246 					 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
247 			return false;
248 		}
249 		fastopenq->rskq_rst_head = req1->dl_next;
250 		fastopenq->qlen--;
251 		spin_unlock(&fastopenq->lock);
252 		reqsk_put(req1);
253 	}
254 	return true;
255 }
256 
257 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
258  * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
259  * cookie request (foc->len == 0).
260  */
261 bool tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
262 		      struct request_sock *req,
263 		      struct tcp_fastopen_cookie *foc,
264 		      struct dst_entry *dst)
265 {
266 	struct tcp_fastopen_cookie valid_foc = { .len = -1 };
267 	bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
268 
269 	if (foc->len == 0) /* Client requests a cookie */
270 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
271 
272 	if (!((sysctl_tcp_fastopen & TFO_SERVER_ENABLE) &&
273 	      (syn_data || foc->len >= 0) &&
274 	      tcp_fastopen_queue_check(sk))) {
275 		foc->len = -1;
276 		return false;
277 	}
278 
279 	if (syn_data && (sysctl_tcp_fastopen & TFO_SERVER_COOKIE_NOT_REQD))
280 		goto fastopen;
281 
282 	if (foc->len >= 0 &&  /* Client presents or requests a cookie */
283 	    tcp_fastopen_cookie_gen(req, skb, &valid_foc) &&
284 	    foc->len == TCP_FASTOPEN_COOKIE_SIZE &&
285 	    foc->len == valid_foc.len &&
286 	    !memcmp(foc->val, valid_foc.val, foc->len)) {
287 		/* Cookie is valid. Create a (full) child socket to accept
288 		 * the data in SYN before returning a SYN-ACK to ack the
289 		 * data. If we fail to create the socket, fall back and
290 		 * ack the ISN only but includes the same cookie.
291 		 *
292 		 * Note: Data-less SYN with valid cookie is allowed to send
293 		 * data in SYN_RECV state.
294 		 */
295 fastopen:
296 		if (tcp_fastopen_create_child(sk, skb, dst, req)) {
297 			foc->len = -1;
298 			NET_INC_STATS_BH(sock_net(sk),
299 					 LINUX_MIB_TCPFASTOPENPASSIVE);
300 			return true;
301 		}
302 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
303 	} else if (foc->len > 0) /* Client presents an invalid cookie */
304 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
305 
306 	valid_foc.exp = foc->exp;
307 	*foc = valid_foc;
308 	return false;
309 }
310 EXPORT_SYMBOL(tcp_try_fastopen);
311