blob: 9771563ab564923e325f5eb5341d7e6ae81a499a [file] [log] [blame]
Jerry Chu10467162012-08-31 12:29:11 +00001#include <linux/err.h>
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00002#include <linux/init.h>
3#include <linux/kernel.h>
Jerry Chu10467162012-08-31 12:29:11 +00004#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>
Yuchung Cheng2100c8d2012-07-19 06:43:05 +000010
Yuchung Cheng0d41cca2013-10-31 09:19:32 -070011int sysctl_tcp_fastopen __read_mostly = TFO_CLIENT_ENABLE;
Jerry Chu10467162012-08-31 12:29:11 +000012
13struct tcp_fastopen_context __rcu *tcp_fastopen_ctx;
14
15static DEFINE_SPINLOCK(tcp_fastopen_ctx_lock);
16
Hannes Frederic Sowa222e83d2013-10-19 21:48:58 +020017void 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
Jerry Chu10467162012-08-31 12:29:11 +000031static 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
39int 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);
51error: 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
Daniel Lee3a19ce02014-05-11 20:22:13 -070075static bool __tcp_fastopen_cookie_gen(const void *path,
76 struct tcp_fastopen_cookie *foc)
Jerry Chu10467162012-08-31 12:29:11 +000077{
Jerry Chu10467162012-08-31 12:29:11 +000078 struct tcp_fastopen_context *ctx;
Daniel Lee3a19ce02014-05-11 20:22:13 -070079 bool ok = false;
Jerry Chu10467162012-08-31 12:29:11 +000080
Hannes Frederic Sowa222e83d2013-10-19 21:48:58 +020081 tcp_fastopen_init_key_once(true);
82
Jerry Chu10467162012-08-31 12:29:11 +000083 rcu_read_lock();
84 ctx = rcu_dereference(tcp_fastopen_ctx);
85 if (ctx) {
Daniel Lee3a19ce02014-05-11 20:22:13 -070086 crypto_cipher_encrypt_one(ctx->tfm, foc->val, path);
Jerry Chu10467162012-08-31 12:29:11 +000087 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
Daniel Lee3a19ce02014-05-11 20:22:13 -070088 ok = true;
Jerry Chu10467162012-08-31 12:29:11 +000089 }
90 rcu_read_unlock();
Daniel Lee3a19ce02014-05-11 20:22:13 -070091 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 */
100static 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;
Jerry Chu10467162012-08-31 12:29:11 +0000127}
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700128
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700129static bool tcp_fastopen_create_child(struct sock *sk,
130 struct sk_buff *skb,
131 struct dst_entry *dst,
132 struct request_sock *req)
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700133{
Dave Jones17846372014-06-16 16:30:36 -0400134 struct tcp_sock *tp;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700135 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700136 struct sock *child;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700137
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);
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700143 if (child == NULL)
144 return false;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700145
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 *
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700189 * 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?
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700192 */
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700193 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1) {
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700194 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);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700199 tp->syn_data_acked = 1;
200 }
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700201 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700202 sk->sk_data_ready(sk);
203 bh_unlock_sock(child);
204 sock_put(child);
205 WARN_ON(req->sk == NULL);
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700206 return true;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700207}
208EXPORT_SYMBOL(tcp_fastopen_create_child);
209
210static 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
Yuchung Cheng89278c92014-05-11 20:22:10 -0700246/* 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 */
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700250bool 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)
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700254{
Yuchung Cheng89278c92014-05-11 20:22:10 -0700255 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
256 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700257
Yuchung Cheng89278c92014-05-11 20:22:10 -0700258 if (!((sysctl_tcp_fastopen & TFO_SERVER_ENABLE) &&
259 (syn_data || foc->len >= 0) &&
260 tcp_fastopen_queue_check(sk))) {
261 foc->len = -1;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700262 return false;
Yuchung Cheng89278c92014-05-11 20:22:10 -0700263 }
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700264
Yuchung Cheng89278c92014-05-11 20:22:10 -0700265 if (syn_data && (sysctl_tcp_fastopen & TFO_SERVER_COOKIE_NOT_REQD))
266 goto fastopen;
267
Daniel Lee3a19ce02014-05-11 20:22:13 -0700268 if (tcp_fastopen_cookie_gen(req, skb, &valid_foc) &&
269 foc->len == TCP_FASTOPEN_COOKIE_SIZE &&
Yuchung Cheng89278c92014-05-11 20:22:10 -0700270 foc->len == valid_foc.len &&
271 !memcmp(foc->val, valid_foc.val, foc->len)) {
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700272 /* 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 */
Yuchung Cheng89278c92014-05-11 20:22:10 -0700280fastopen:
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700281 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 }
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700287 }
288
Yuchung Cheng89278c92014-05-11 20:22:10 -0700289 NET_INC_STATS_BH(sock_net(sk), foc->len ?
290 LINUX_MIB_TCPFASTOPENPASSIVEFAIL :
291 LINUX_MIB_TCPFASTOPENCOOKIEREQD);
292 *foc = valid_foc;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700293 return false;
294}
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700295EXPORT_SYMBOL(tcp_try_fastopen);