blob: 93396bf7b475f3972d247d282faa406d962696dd [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
81 rcu_read_lock();
82 ctx = rcu_dereference(tcp_fastopen_ctx);
83 if (ctx) {
Daniel Lee3a19ce02014-05-11 20:22:13 -070084 crypto_cipher_encrypt_one(ctx->tfm, foc->val, path);
Jerry Chu10467162012-08-31 12:29:11 +000085 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
Daniel Lee3a19ce02014-05-11 20:22:13 -070086 ok = true;
Jerry Chu10467162012-08-31 12:29:11 +000087 }
88 rcu_read_unlock();
Daniel Lee3a19ce02014-05-11 20:22:13 -070089 return ok;
90}
91
92/* Generate the fastopen cookie by doing aes128 encryption on both
93 * the source and destination addresses. Pad 0s for IPv4 or IPv4-mapped-IPv6
94 * addresses. For the longer IPv6 addresses use CBC-MAC.
95 *
96 * XXX (TFO) - refactor when TCP_FASTOPEN_COOKIE_SIZE != AES_BLOCK_SIZE.
97 */
98static bool tcp_fastopen_cookie_gen(struct request_sock *req,
99 struct sk_buff *syn,
100 struct tcp_fastopen_cookie *foc)
101{
102 if (req->rsk_ops->family == AF_INET) {
103 const struct iphdr *iph = ip_hdr(syn);
104
105 __be32 path[4] = { iph->saddr, iph->daddr, 0, 0 };
106 return __tcp_fastopen_cookie_gen(path, foc);
107 }
108
109#if IS_ENABLED(CONFIG_IPV6)
110 if (req->rsk_ops->family == AF_INET6) {
111 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
112 struct tcp_fastopen_cookie tmp;
113
114 if (__tcp_fastopen_cookie_gen(&ip6h->saddr, &tmp)) {
115 struct in6_addr *buf = (struct in6_addr *) tmp.val;
Li RongQing41c91992014-09-29 15:04:37 +0800116 int i;
Daniel Lee3a19ce02014-05-11 20:22:13 -0700117
118 for (i = 0; i < 4; i++)
119 buf->s6_addr32[i] ^= ip6h->daddr.s6_addr32[i];
120 return __tcp_fastopen_cookie_gen(buf, foc);
121 }
122 }
123#endif
124 return false;
Jerry Chu10467162012-08-31 12:29:11 +0000125}
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700126
Eric Dumazet7c85af82015-09-24 17:16:05 -0700127static struct sock *tcp_fastopen_create_child(struct sock *sk,
128 struct sk_buff *skb,
129 struct dst_entry *dst,
130 struct request_sock *req)
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700131{
Dave Jones17846372014-06-16 16:30:36 -0400132 struct tcp_sock *tp;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700133 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700134 struct sock *child;
Eric Dumazetba34e6d2015-02-13 04:47:12 -0800135 u32 end_seq;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700136
137 req->num_retrans = 0;
138 req->num_timeout = 0;
139 req->sk = NULL;
140
141 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
Ian Morris51456b22015-04-03 09:17:26 +0100142 if (!child)
Eric Dumazet7c85af82015-09-24 17:16:05 -0700143 return NULL;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700144
Eric Dumazet0536fcc2015-09-29 07:42:52 -0700145 spin_lock(&queue->fastopenq.lock);
146 queue->fastopenq.qlen++;
147 spin_unlock(&queue->fastopenq.lock);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700148
149 /* Initialize the child socket. Have to fix some values to take
150 * into account the child is a Fast Open socket and is created
151 * only out of the bits carried in the SYN packet.
152 */
153 tp = tcp_sk(child);
154
155 tp->fastopen_rsk = req;
Eric Dumazet9439ce02015-03-17 18:32:29 -0700156 tcp_rsk(req)->tfo_listener = true;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700157
158 /* RFC1323: The window in SYN & SYN/ACK segments is never
159 * scaled. So correct it appropriately.
160 */
161 tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
162
163 /* Activate the retrans timer so that SYNACK can be retransmitted.
Eric Dumazetca6fb062015-10-02 11:43:35 -0700164 * The request socket is not added to the ehash
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700165 * because it's been added to the accept queue directly.
166 */
167 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
168 TCP_TIMEOUT_INIT, TCP_RTO_MAX);
169
Eric Dumazetca6fb062015-10-02 11:43:35 -0700170 atomic_set(&req->rsk_refcnt, 2);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700171
172 /* Now finish processing the fastopen child socket. */
173 inet_csk(child)->icsk_af_ops->rebuild_header(child);
174 tcp_init_congestion_control(child);
175 tcp_mtup_init(child);
176 tcp_init_metrics(child);
177 tcp_init_buffer_space(child);
178
Eric Dumazet7656d842015-10-04 21:08:07 -0700179 /* Queue the data carried in the SYN packet.
180 * We used to play tricky games with skb_get().
181 * With lockless listener, it is a dead end.
182 * Do not think about it.
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700183 *
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700184 * XXX (TFO) - we honor a zero-payload TFO request for now,
185 * (any reason not to?) but no need to queue the skb since
186 * there is no data. How about SYN+FIN?
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700187 */
Eric Dumazetba34e6d2015-02-13 04:47:12 -0800188 end_seq = TCP_SKB_CB(skb)->end_seq;
189 if (end_seq != TCP_SKB_CB(skb)->seq + 1) {
Eric Dumazet7656d842015-10-04 21:08:07 -0700190 struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
Eric Dumazetba34e6d2015-02-13 04:47:12 -0800191
192 if (likely(skb2)) {
193 skb_dst_drop(skb2);
194 __skb_pull(skb2, tcp_hdrlen(skb));
195 skb_set_owner_r(skb2, child);
196 __skb_queue_tail(&child->sk_receive_queue, skb2);
197 tp->syn_data_acked = 1;
Eric Dumazetd6549762015-05-21 21:51:19 -0700198
199 /* u64_stats_update_begin(&tp->syncp) not needed here,
200 * as we certainly are not changing upper 32bit value (0)
201 */
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -0700202 tp->bytes_received = end_seq - TCP_SKB_CB(skb)->seq - 1;
Eric Dumazetba34e6d2015-02-13 04:47:12 -0800203 } else {
204 end_seq = TCP_SKB_CB(skb)->seq + 1;
205 }
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700206 }
Eric Dumazetba34e6d2015-02-13 04:47:12 -0800207 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt = end_seq;
Eric Dumazet7656d842015-10-04 21:08:07 -0700208 /* tcp_conn_request() is sending the SYNACK,
209 * and queues the child into listener accept queue.
Eric Dumazet7c85af82015-09-24 17:16:05 -0700210 */
Eric Dumazet7c85af82015-09-24 17:16:05 -0700211 return child;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700212}
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700213
214static bool tcp_fastopen_queue_check(struct sock *sk)
215{
216 struct fastopen_queue *fastopenq;
217
218 /* Make sure the listener has enabled fastopen, and we don't
219 * exceed the max # of pending TFO requests allowed before trying
220 * to validating the cookie in order to avoid burning CPU cycles
221 * unnecessarily.
222 *
223 * XXX (TFO) - The implication of checking the max_qlen before
224 * processing a cookie request is that clients can't differentiate
225 * between qlen overflow causing Fast Open to be disabled
226 * temporarily vs a server not supporting Fast Open at all.
227 */
Eric Dumazet0536fcc2015-09-29 07:42:52 -0700228 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
229 if (fastopenq->max_qlen == 0)
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700230 return false;
231
232 if (fastopenq->qlen >= fastopenq->max_qlen) {
233 struct request_sock *req1;
234 spin_lock(&fastopenq->lock);
235 req1 = fastopenq->rskq_rst_head;
Eric Dumazetfa76ce732015-03-19 19:04:20 -0700236 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700237 spin_unlock(&fastopenq->lock);
238 NET_INC_STATS_BH(sock_net(sk),
239 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
240 return false;
241 }
242 fastopenq->rskq_rst_head = req1->dl_next;
243 fastopenq->qlen--;
244 spin_unlock(&fastopenq->lock);
Eric Dumazet13854e52015-03-15 21:12:16 -0700245 reqsk_put(req1);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700246 }
247 return true;
248}
249
Yuchung Cheng89278c92014-05-11 20:22:10 -0700250/* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
251 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
252 * cookie request (foc->len == 0).
253 */
Eric Dumazet7c85af82015-09-24 17:16:05 -0700254struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
255 struct request_sock *req,
256 struct tcp_fastopen_cookie *foc,
257 struct dst_entry *dst)
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700258{
Yuchung Cheng89278c92014-05-11 20:22:10 -0700259 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
260 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
Eric Dumazet7c85af82015-09-24 17:16:05 -0700261 struct sock *child;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700262
Yuchung Cheng531c94a2015-02-09 12:35:23 -0800263 if (foc->len == 0) /* Client requests a cookie */
264 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
265
Yuchung Cheng89278c92014-05-11 20:22:10 -0700266 if (!((sysctl_tcp_fastopen & TFO_SERVER_ENABLE) &&
267 (syn_data || foc->len >= 0) &&
268 tcp_fastopen_queue_check(sk))) {
269 foc->len = -1;
Eric Dumazet7c85af82015-09-24 17:16:05 -0700270 return NULL;
Yuchung Cheng89278c92014-05-11 20:22:10 -0700271 }
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700272
Yuchung Cheng89278c92014-05-11 20:22:10 -0700273 if (syn_data && (sysctl_tcp_fastopen & TFO_SERVER_COOKIE_NOT_REQD))
274 goto fastopen;
275
Yuchung Cheng531c94a2015-02-09 12:35:23 -0800276 if (foc->len >= 0 && /* Client presents or requests a cookie */
277 tcp_fastopen_cookie_gen(req, skb, &valid_foc) &&
Daniel Lee3a19ce02014-05-11 20:22:13 -0700278 foc->len == TCP_FASTOPEN_COOKIE_SIZE &&
Yuchung Cheng89278c92014-05-11 20:22:10 -0700279 foc->len == valid_foc.len &&
280 !memcmp(foc->val, valid_foc.val, foc->len)) {
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700281 /* Cookie is valid. Create a (full) child socket to accept
282 * the data in SYN before returning a SYN-ACK to ack the
283 * data. If we fail to create the socket, fall back and
284 * ack the ISN only but includes the same cookie.
285 *
286 * Note: Data-less SYN with valid cookie is allowed to send
287 * data in SYN_RECV state.
288 */
Yuchung Cheng89278c92014-05-11 20:22:10 -0700289fastopen:
Eric Dumazet7c85af82015-09-24 17:16:05 -0700290 child = tcp_fastopen_create_child(sk, skb, dst, req);
291 if (child) {
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700292 foc->len = -1;
293 NET_INC_STATS_BH(sock_net(sk),
294 LINUX_MIB_TCPFASTOPENPASSIVE);
Eric Dumazet7c85af82015-09-24 17:16:05 -0700295 return child;
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700296 }
Yuchung Cheng531c94a2015-02-09 12:35:23 -0800297 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
298 } else if (foc->len > 0) /* Client presents an invalid cookie */
299 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700300
Daniel Lee7f9b8382015-04-06 14:37:26 -0700301 valid_foc.exp = foc->exp;
Yuchung Cheng89278c92014-05-11 20:22:10 -0700302 *foc = valid_foc;
Eric Dumazet7c85af82015-09-24 17:16:05 -0700303 return NULL;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700304}