blob: 45cc7f1ca29618e3ac1066cb49e7d6dc90e1c64d [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00002#include <linux/kernel.h>
Jerry Chu10467162012-08-31 12:29:11 +00003#include <linux/tcp.h>
4#include <linux/rcupdate.h>
Jerry Chu10467162012-08-31 12:29:11 +00005#include <net/tcp.h>
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00006
Haishuang Yan43713842017-09-27 11:35:42 +08007void tcp_fastopen_init_key_once(struct net *net)
Hannes Frederic Sowa222e83d2013-10-19 21:48:58 +02008{
Haishuang Yan43713842017-09-27 11:35:42 +08009 u8 key[TCP_FASTOPEN_KEY_LENGTH];
10 struct tcp_fastopen_context *ctxt;
11
12 rcu_read_lock();
13 ctxt = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
14 if (ctxt) {
15 rcu_read_unlock();
16 return;
17 }
18 rcu_read_unlock();
Hannes Frederic Sowa222e83d2013-10-19 21:48:58 +020019
20 /* tcp_fastopen_reset_cipher publishes the new context
21 * atomically, so we allow this race happening here.
22 *
23 * All call sites of tcp_fastopen_cookie_gen also check
24 * for a valid cookie, so this is an acceptable risk.
25 */
Haishuang Yan43713842017-09-27 11:35:42 +080026 get_random_bytes(key, sizeof(key));
Ard Biesheuvel438ac882019-06-19 23:46:28 +020027 tcp_fastopen_reset_cipher(net, NULL, key, NULL);
Hannes Frederic Sowa222e83d2013-10-19 21:48:58 +020028}
29
Jerry Chu10467162012-08-31 12:29:11 +000030static void tcp_fastopen_ctx_free(struct rcu_head *head)
31{
32 struct tcp_fastopen_context *ctx =
33 container_of(head, struct tcp_fastopen_context, rcu);
Jason Baron9092a762019-05-29 12:33:57 -040034
Waiman Long453431a2020-08-06 23:18:13 -070035 kfree_sensitive(ctx);
Jerry Chu10467162012-08-31 12:29:11 +000036}
37
Yuchung Cheng1fba70e2017-10-18 11:22:51 -070038void tcp_fastopen_destroy_cipher(struct sock *sk)
39{
40 struct tcp_fastopen_context *ctx;
41
42 ctx = rcu_dereference_protected(
43 inet_csk(sk)->icsk_accept_queue.fastopenq.ctx, 1);
44 if (ctx)
45 call_rcu(&ctx->rcu, tcp_fastopen_ctx_free);
46}
47
Haishuang Yan43713842017-09-27 11:35:42 +080048void tcp_fastopen_ctx_destroy(struct net *net)
49{
50 struct tcp_fastopen_context *ctxt;
51
Eric Dumazete93abb82021-07-19 03:11:07 -070052 ctxt = xchg((__force struct tcp_fastopen_context **)&net->ipv4.tcp_fastopen_ctx, NULL);
Haishuang Yan43713842017-09-27 11:35:42 +080053
54 if (ctxt)
55 call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free);
56}
57
Yuchung Cheng1fba70e2017-10-18 11:22:51 -070058int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
Ard Biesheuvel438ac882019-06-19 23:46:28 +020059 void *primary_key, void *backup_key)
Jerry Chu10467162012-08-31 12:29:11 +000060{
Jerry Chu10467162012-08-31 12:29:11 +000061 struct tcp_fastopen_context *ctx, *octx;
Yuchung Cheng1fba70e2017-10-18 11:22:51 -070062 struct fastopen_queue *q;
Jason Baron9092a762019-05-29 12:33:57 -040063 int err = 0;
Jerry Chu10467162012-08-31 12:29:11 +000064
Ard Biesheuvelc681eda2019-06-17 10:09:33 +020065 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
66 if (!ctx) {
67 err = -ENOMEM;
Jason Baron9092a762019-05-29 12:33:57 -040068 goto out;
Jerry Chu10467162012-08-31 12:29:11 +000069 }
Ard Biesheuvelc681eda2019-06-17 10:09:33 +020070
Ard Biesheuvel438ac882019-06-19 23:46:28 +020071 ctx->key[0].key[0] = get_unaligned_le64(primary_key);
72 ctx->key[0].key[1] = get_unaligned_le64(primary_key + 8);
Ard Biesheuvelc681eda2019-06-17 10:09:33 +020073 if (backup_key) {
Ard Biesheuvel438ac882019-06-19 23:46:28 +020074 ctx->key[1].key[0] = get_unaligned_le64(backup_key);
75 ctx->key[1].key[1] = get_unaligned_le64(backup_key + 8);
Ard Biesheuvelc681eda2019-06-17 10:09:33 +020076 ctx->num = 2;
77 } else {
78 ctx->num = 1;
79 }
80
Yuchung Cheng1fba70e2017-10-18 11:22:51 -070081 if (sk) {
82 q = &inet_csk(sk)->icsk_accept_queue.fastopenq;
Eric Dumazete93abb82021-07-19 03:11:07 -070083 octx = xchg((__force struct tcp_fastopen_context **)&q->ctx, ctx);
Yuchung Cheng1fba70e2017-10-18 11:22:51 -070084 } else {
Eric Dumazete93abb82021-07-19 03:11:07 -070085 octx = xchg((__force struct tcp_fastopen_context **)&net->ipv4.tcp_fastopen_ctx, ctx);
Yuchung Cheng1fba70e2017-10-18 11:22:51 -070086 }
Jerry Chu10467162012-08-31 12:29:11 +000087
88 if (octx)
89 call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
Jason Baron9092a762019-05-29 12:33:57 -040090out:
Jerry Chu10467162012-08-31 12:29:11 +000091 return err;
92}
93
Jason Baronf19008e2020-08-10 13:38:39 -040094int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
95 u64 *key)
96{
97 struct tcp_fastopen_context *ctx;
98 int n_keys = 0, i;
99
100 rcu_read_lock();
101 if (icsk)
102 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
103 else
104 ctx = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
105 if (ctx) {
106 n_keys = tcp_fastopen_context_len(ctx);
107 for (i = 0; i < n_keys; i++) {
108 put_unaligned_le64(ctx->key[i].key[0], key + (i * 2));
109 put_unaligned_le64(ctx->key[i].key[1], key + (i * 2) + 1);
110 }
111 }
112 rcu_read_unlock();
113
114 return n_keys;
115}
116
Christoph Paasch483642e2019-05-29 12:33:56 -0400117static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req,
118 struct sk_buff *syn,
Ard Biesheuvel438ac882019-06-19 23:46:28 +0200119 const siphash_key_t *key,
Christoph Paasch483642e2019-05-29 12:33:56 -0400120 struct tcp_fastopen_cookie *foc)
Jerry Chu10467162012-08-31 12:29:11 +0000121{
Ard Biesheuvelc681eda2019-06-17 10:09:33 +0200122 BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64));
123
Christoph Paasch483642e2019-05-29 12:33:56 -0400124 if (req->rsk_ops->family == AF_INET) {
125 const struct iphdr *iph = ip_hdr(syn);
Jerry Chu10467162012-08-31 12:29:11 +0000126
Ard Biesheuvel438ac882019-06-19 23:46:28 +0200127 foc->val[0] = cpu_to_le64(siphash(&iph->saddr,
128 sizeof(iph->saddr) +
129 sizeof(iph->daddr),
130 key));
Jerry Chu10467162012-08-31 12:29:11 +0000131 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
Christoph Paasch483642e2019-05-29 12:33:56 -0400132 return true;
Jerry Chu10467162012-08-31 12:29:11 +0000133 }
Christoph Paasch483642e2019-05-29 12:33:56 -0400134#if IS_ENABLED(CONFIG_IPV6)
135 if (req->rsk_ops->family == AF_INET6) {
136 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
Christoph Paasch483642e2019-05-29 12:33:56 -0400137
Ard Biesheuvel438ac882019-06-19 23:46:28 +0200138 foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr,
139 sizeof(ip6h->saddr) +
140 sizeof(ip6h->daddr),
141 key));
Christoph Paasch483642e2019-05-29 12:33:56 -0400142 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
143 return true;
144 }
145#endif
146 return false;
Daniel Lee3a19ce02014-05-11 20:22:13 -0700147}
148
Ard Biesheuvelc681eda2019-06-17 10:09:33 +0200149/* Generate the fastopen cookie by applying SipHash to both the source and
150 * destination addresses.
Daniel Lee3a19ce02014-05-11 20:22:13 -0700151 */
Jason Baron9092a762019-05-29 12:33:57 -0400152static void tcp_fastopen_cookie_gen(struct sock *sk,
Haishuang Yan43713842017-09-27 11:35:42 +0800153 struct request_sock *req,
Daniel Lee3a19ce02014-05-11 20:22:13 -0700154 struct sk_buff *syn,
155 struct tcp_fastopen_cookie *foc)
156{
Christoph Paasch483642e2019-05-29 12:33:56 -0400157 struct tcp_fastopen_context *ctx;
Daniel Lee3a19ce02014-05-11 20:22:13 -0700158
Christoph Paasch483642e2019-05-29 12:33:56 -0400159 rcu_read_lock();
Jason Baron9092a762019-05-29 12:33:57 -0400160 ctx = tcp_fastopen_get_ctx(sk);
Christoph Paasch483642e2019-05-29 12:33:56 -0400161 if (ctx)
Ard Biesheuvel438ac882019-06-19 23:46:28 +0200162 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc);
Christoph Paasch483642e2019-05-29 12:33:56 -0400163 rcu_read_unlock();
Jerry Chu10467162012-08-31 12:29:11 +0000164}
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700165
Eric Dumazet61d2bca2016-02-01 21:03:07 -0800166/* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
167 * queue this additional data / FIN.
168 */
169void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
170{
171 struct tcp_sock *tp = tcp_sk(sk);
172
173 if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
174 return;
175
176 skb = skb_clone(skb, GFP_ATOMIC);
177 if (!skb)
178 return;
179
180 skb_dst_drop(skb);
Martin KaFai Laua44d6ea2016-03-14 10:52:15 -0700181 /* segs_in has been initialized to 1 in tcp_create_openreq_child().
182 * Hence, reset segs_in to 0 before calling tcp_segs_in()
183 * to avoid double counting. Also, tcp_segs_in() expects
184 * skb->len to include the tcp_hdrlen. Hence, it should
185 * be called before __skb_pull().
186 */
187 tp->segs_in = 0;
188 tcp_segs_in(tp, skb);
Eric Dumazet61d2bca2016-02-01 21:03:07 -0800189 __skb_pull(skb, tcp_hdrlen(skb));
Eric Dumazet76061f62016-09-07 08:34:11 -0700190 sk_forced_mem_schedule(sk, skb->truesize);
Eric Dumazet61d2bca2016-02-01 21:03:07 -0800191 skb_set_owner_r(skb, sk);
192
Eric Dumazet9d691532016-02-01 21:03:08 -0800193 TCP_SKB_CB(skb)->seq++;
194 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
195
Eric Dumazet61d2bca2016-02-01 21:03:07 -0800196 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
197 __skb_queue_tail(&sk->sk_receive_queue, skb);
198 tp->syn_data_acked = 1;
199
200 /* u64_stats_update_begin(&tp->syncp) not needed here,
201 * as we certainly are not changing upper 32bit value (0)
202 */
203 tp->bytes_received = skb->len;
Eric Dumazete3e17b72016-02-06 11:16:28 -0800204
205 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
206 tcp_fin(sk);
Eric Dumazet61d2bca2016-02-01 21:03:07 -0800207}
208
Jason Baron9092a762019-05-29 12:33:57 -0400209/* returns 0 - no key match, 1 for primary, 2 for backup */
210static int tcp_fastopen_cookie_gen_check(struct sock *sk,
211 struct request_sock *req,
212 struct sk_buff *syn,
213 struct tcp_fastopen_cookie *orig,
214 struct tcp_fastopen_cookie *valid_foc)
215{
216 struct tcp_fastopen_cookie search_foc = { .len = -1 };
217 struct tcp_fastopen_cookie *foc = valid_foc;
218 struct tcp_fastopen_context *ctx;
219 int i, ret = 0;
220
221 rcu_read_lock();
222 ctx = tcp_fastopen_get_ctx(sk);
223 if (!ctx)
224 goto out;
225 for (i = 0; i < tcp_fastopen_context_len(ctx); i++) {
Ard Biesheuvel438ac882019-06-19 23:46:28 +0200226 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[i], foc);
Jason Baron9092a762019-05-29 12:33:57 -0400227 if (tcp_fastopen_cookie_match(foc, orig)) {
228 ret = i + 1;
229 goto out;
230 }
231 foc = &search_foc;
232 }
233out:
234 rcu_read_unlock();
235 return ret;
236}
237
Eric Dumazet7c85af82015-09-24 17:16:05 -0700238static struct sock *tcp_fastopen_create_child(struct sock *sk,
239 struct sk_buff *skb,
Eric Dumazet7c85af82015-09-24 17:16:05 -0700240 struct request_sock *req)
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700241{
Dave Jones17846372014-06-16 16:30:36 -0400242 struct tcp_sock *tp;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700243 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700244 struct sock *child;
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700245 bool own_req;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700246
Eric Dumazet5e0724d2015-10-22 08:20:46 -0700247 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
248 NULL, &own_req);
Ian Morris51456b22015-04-03 09:17:26 +0100249 if (!child)
Eric Dumazet7c85af82015-09-24 17:16:05 -0700250 return NULL;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700251
Eric Dumazet0536fcc2015-09-29 07:42:52 -0700252 spin_lock(&queue->fastopenq.lock);
253 queue->fastopenq.qlen++;
254 spin_unlock(&queue->fastopenq.lock);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700255
256 /* Initialize the child socket. Have to fix some values to take
257 * into account the child is a Fast Open socket and is created
258 * only out of the bits carried in the SYN packet.
259 */
260 tp = tcp_sk(child);
261
Eric Dumazetd983ea62019-10-10 20:17:38 -0700262 rcu_assign_pointer(tp->fastopen_rsk, req);
Eric Dumazet9439ce02015-03-17 18:32:29 -0700263 tcp_rsk(req)->tfo_listener = true;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700264
265 /* RFC1323: The window in SYN & SYN/ACK segments is never
266 * scaled. So correct it appropriately.
267 */
268 tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
Alexey Kodanev0dbd7ff2017-01-19 16:36:39 +0300269 tp->max_window = tp->snd_wnd;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700270
271 /* Activate the retrans timer so that SYNACK can be retransmitted.
Eric Dumazetca6fb062015-10-02 11:43:35 -0700272 * The request socket is not added to the ehash
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700273 * because it's been added to the accept queue directly.
274 */
Jie Meng8ea731d2022-08-15 13:29:00 -0700275 req->timeout = tcp_timeout_init(child);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700276 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
Jie Meng8ea731d2022-08-15 13:29:00 -0700277 req->timeout, TCP_RTO_MAX);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700278
Reshetova, Elena41c6d652017-06-30 13:08:01 +0300279 refcount_set(&req->rsk_refcnt, 2);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700280
281 /* Now finish processing the fastopen child socket. */
Martin KaFai Lau72be0fe2020-08-20 12:00:39 -0700282 tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, skb);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700283
Eric Dumazet61d2bca2016-02-01 21:03:07 -0800284 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Eric Dumazetba34e6d2015-02-13 04:47:12 -0800285
Eric Dumazet61d2bca2016-02-01 21:03:07 -0800286 tcp_fastopen_add_skb(child, skb);
Eric Dumazetd6549762015-05-21 21:51:19 -0700287
Eric Dumazet61d2bca2016-02-01 21:03:07 -0800288 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
Neal Cardwell28b346c2016-08-30 11:55:23 -0400289 tp->rcv_wup = tp->rcv_nxt;
Eric Dumazet7656d842015-10-04 21:08:07 -0700290 /* tcp_conn_request() is sending the SYNACK,
291 * and queues the child into listener accept queue.
Eric Dumazet7c85af82015-09-24 17:16:05 -0700292 */
Eric Dumazet7c85af82015-09-24 17:16:05 -0700293 return child;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700294}
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700295
296static bool tcp_fastopen_queue_check(struct sock *sk)
297{
298 struct fastopen_queue *fastopenq;
299
300 /* Make sure the listener has enabled fastopen, and we don't
301 * exceed the max # of pending TFO requests allowed before trying
302 * to validating the cookie in order to avoid burning CPU cycles
303 * unnecessarily.
304 *
305 * XXX (TFO) - The implication of checking the max_qlen before
306 * processing a cookie request is that clients can't differentiate
307 * between qlen overflow causing Fast Open to be disabled
308 * temporarily vs a server not supporting Fast Open at all.
309 */
Eric Dumazet0536fcc2015-09-29 07:42:52 -0700310 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
311 if (fastopenq->max_qlen == 0)
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700312 return false;
313
314 if (fastopenq->qlen >= fastopenq->max_qlen) {
315 struct request_sock *req1;
316 spin_lock(&fastopenq->lock);
317 req1 = fastopenq->rskq_rst_head;
Eric Dumazetfa76ce732015-03-19 19:04:20 -0700318 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
Eric Dumazet02a1d6e2016-04-27 16:44:39 -0700319 __NET_INC_STATS(sock_net(sk),
320 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
Eric Dumazetc10d9312016-04-29 14:16:47 -0700321 spin_unlock(&fastopenq->lock);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700322 return false;
323 }
324 fastopenq->rskq_rst_head = req1->dl_next;
325 fastopenq->qlen--;
326 spin_unlock(&fastopenq->lock);
Eric Dumazet13854e52015-03-15 21:12:16 -0700327 reqsk_put(req1);
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700328 }
329 return true;
330}
331
Christoph Paasch71c02372017-10-23 13:22:23 -0700332static bool tcp_fastopen_no_cookie(const struct sock *sk,
333 const struct dst_entry *dst,
334 int flag)
335{
Kuniyuki Iwashima5a542132022-07-15 10:17:54 -0700336 return (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) & flag) ||
Christoph Paasch71c02372017-10-23 13:22:23 -0700337 tcp_sk(sk)->fastopen_no_cookie ||
338 (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
339}
340
Yuchung Cheng89278c92014-05-11 20:22:10 -0700341/* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
342 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
343 * cookie request (foc->len == 0).
344 */
Eric Dumazet7c85af82015-09-24 17:16:05 -0700345struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
346 struct request_sock *req,
Christoph Paasch71c02372017-10-23 13:22:23 -0700347 struct tcp_fastopen_cookie *foc,
348 const struct dst_entry *dst)
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700349{
Yuchung Cheng89278c92014-05-11 20:22:10 -0700350 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
Kuniyuki Iwashima5a542132022-07-15 10:17:54 -0700351 int tcp_fastopen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen);
Haishuang Yane1cfcbe2017-09-27 11:35:40 +0800352 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
Eric Dumazet7c85af82015-09-24 17:16:05 -0700353 struct sock *child;
Jason Baron9092a762019-05-29 12:33:57 -0400354 int ret = 0;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700355
Yuchung Cheng531c94a2015-02-09 12:35:23 -0800356 if (foc->len == 0) /* Client requests a cookie */
Eric Dumazetc10d9312016-04-29 14:16:47 -0700357 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
Yuchung Cheng531c94a2015-02-09 12:35:23 -0800358
Haishuang Yane1cfcbe2017-09-27 11:35:40 +0800359 if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
Yuchung Cheng89278c92014-05-11 20:22:10 -0700360 (syn_data || foc->len >= 0) &&
361 tcp_fastopen_queue_check(sk))) {
362 foc->len = -1;
Eric Dumazet7c85af82015-09-24 17:16:05 -0700363 return NULL;
Yuchung Cheng89278c92014-05-11 20:22:10 -0700364 }
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700365
Luke Hsiaoe3faa492021-08-16 20:51:06 +0000366 if (tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
Yuchung Cheng89278c92014-05-11 20:22:10 -0700367 goto fastopen;
368
Jason Baron9092a762019-05-29 12:33:57 -0400369 if (foc->len == 0) {
370 /* Client requests a cookie. */
371 tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc);
372 } else if (foc->len > 0) {
373 ret = tcp_fastopen_cookie_gen_check(sk, req, skb, foc,
374 &valid_foc);
375 if (!ret) {
Eric Dumazetc10d9312016-04-29 14:16:47 -0700376 NET_INC_STATS(sock_net(sk),
Jason Baron9092a762019-05-29 12:33:57 -0400377 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
378 } else {
379 /* Cookie is valid. Create a (full) child socket to
380 * accept the data in SYN before returning a SYN-ACK to
381 * ack the data. If we fail to create the socket, fall
382 * back and ack the ISN only but includes the same
383 * cookie.
384 *
385 * Note: Data-less SYN with valid cookie is allowed to
386 * send data in SYN_RECV state.
387 */
388fastopen:
389 child = tcp_fastopen_create_child(sk, skb, req);
390 if (child) {
391 if (ret == 2) {
392 valid_foc.exp = foc->exp;
393 *foc = valid_foc;
394 NET_INC_STATS(sock_net(sk),
395 LINUX_MIB_TCPFASTOPENPASSIVEALTKEY);
396 } else {
397 foc->len = -1;
398 }
399 NET_INC_STATS(sock_net(sk),
400 LINUX_MIB_TCPFASTOPENPASSIVE);
401 return child;
402 }
403 NET_INC_STATS(sock_net(sk),
404 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
Yuchung Cheng843f4a52014-05-11 20:22:11 -0700405 }
Jason Baron9092a762019-05-29 12:33:57 -0400406 }
Daniel Lee7f9b8382015-04-06 14:37:26 -0700407 valid_foc.exp = foc->exp;
Yuchung Cheng89278c92014-05-11 20:22:10 -0700408 *foc = valid_foc;
Eric Dumazet7c85af82015-09-24 17:16:05 -0700409 return NULL;
Yuchung Cheng5b7ed082014-05-11 20:22:09 -0700410}
Wei Wang065263f2017-01-23 10:59:20 -0800411
412bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
413 struct tcp_fastopen_cookie *cookie)
414{
Christoph Paasch71c02372017-10-23 13:22:23 -0700415 const struct dst_entry *dst;
Wei Wang065263f2017-01-23 10:59:20 -0800416
Yuchung Cheng72685862017-12-12 13:10:40 -0800417 tcp_fastopen_cache_get(sk, mss, cookie);
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700418
419 /* Firewall blackhole issue check */
420 if (tcp_fastopen_active_should_disable(sk)) {
421 cookie->len = -1;
422 return false;
423 }
424
Christoph Paasch71c02372017-10-23 13:22:23 -0700425 dst = __sk_dst_get(sk);
426
427 if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
Wei Wang065263f2017-01-23 10:59:20 -0800428 cookie->len = -1;
429 return true;
430 }
Jason Baron48027472019-10-23 11:09:26 -0400431 if (cookie->len > 0)
432 return true;
433 tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE;
434 return false;
Wei Wang065263f2017-01-23 10:59:20 -0800435}
Wei Wang19f6d3f2017-01-23 10:59:22 -0800436
437/* This function checks if we want to defer sending SYN until the first
438 * write(). We defer under the following conditions:
439 * 1. fastopen_connect sockopt is set
440 * 2. we have a valid cookie
441 * Return value: return true if we want to defer until application writes data
442 * return false if we want to send out SYN immediately
443 */
444bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
445{
446 struct tcp_fastopen_cookie cookie = { .len = 0 };
447 struct tcp_sock *tp = tcp_sk(sk);
448 u16 mss;
449
450 if (tp->fastopen_connect && !tp->fastopen_req) {
451 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
452 inet_sk(sk)->defer_connect = 1;
453 return true;
454 }
455
456 /* Alloc fastopen_req in order for FO option to be included
457 * in SYN
458 */
459 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
460 sk->sk_allocation);
461 if (tp->fastopen_req)
462 tp->fastopen_req->cookie = cookie;
463 else
464 *err = -ENOBUFS;
465 }
466 return false;
467}
468EXPORT_SYMBOL(tcp_fastopen_defer_connect);
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700469
470/*
471 * The following code block is to deal with middle box issues with TFO:
472 * Middlebox firewall issues can potentially cause server's data being
473 * blackholed after a successful 3WHS using TFO.
474 * The proposed solution is to disable active TFO globally under the
475 * following circumstances:
476 * 1. client side TFO socket receives out of order FIN
477 * 2. client side TFO socket receives out of order RST
Yuchung Cheng72685862017-12-12 13:10:40 -0800478 * 3. client side TFO socket has timed out three times consecutively during
479 * or after handshake
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700480 * We disable active side TFO globally for 1hr at first. Then if it
481 * happens again, we disable it for 2h, then 4h, 8h, ...
482 * And we reset the timeout back to 1hr when we see a successful active
483 * TFO connection with data exchanges.
484 */
485
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700486/* Disable active TFO and record current jiffies and
487 * tfo_active_disable_times
488 */
Wei Wang46c2fa32017-04-20 14:45:47 -0700489void tcp_fastopen_active_disable(struct sock *sk)
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700490{
Haishuang Yan3733be12017-09-27 11:35:43 +0800491 struct net *net = sock_net(sk);
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700492
Kuniyuki Iwashima021266e2022-07-15 10:17:55 -0700493 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout))
Wei Wang213ad732021-07-21 10:27:38 -0700494 return;
495
Eric Dumazet6f20c8a2021-07-19 02:12:18 -0700496 /* Paired with READ_ONCE() in tcp_fastopen_active_should_disable() */
497 WRITE_ONCE(net->ipv4.tfo_active_disable_stamp, jiffies);
498
499 /* Paired with smp_rmb() in tcp_fastopen_active_should_disable().
500 * We want net->ipv4.tfo_active_disable_stamp to be updated first.
501 */
502 smp_mb__before_atomic();
Haishuang Yan3733be12017-09-27 11:35:43 +0800503 atomic_inc(&net->ipv4.tfo_active_disable_times);
Eric Dumazet6f20c8a2021-07-19 02:12:18 -0700504
Haishuang Yan3733be12017-09-27 11:35:43 +0800505 NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700506}
507
508/* Calculate timeout for tfo active disable
509 * Return true if we are still in the active TFO disable period
510 * Return false if timeout already expired and we should use active TFO
511 */
512bool tcp_fastopen_active_should_disable(struct sock *sk)
513{
Kuniyuki Iwashima021266e2022-07-15 10:17:55 -0700514 unsigned int tfo_bh_timeout =
515 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout);
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700516 unsigned long timeout;
Wei Wang213ad732021-07-21 10:27:38 -0700517 int tfo_da_times;
Haishuang Yan3733be12017-09-27 11:35:43 +0800518 int multiplier;
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700519
Wei Wang213ad732021-07-21 10:27:38 -0700520 if (!tfo_bh_timeout)
521 return false;
522
523 tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700524 if (!tfo_da_times)
525 return false;
526
Eric Dumazet6f20c8a2021-07-19 02:12:18 -0700527 /* Paired with smp_mb__before_atomic() in tcp_fastopen_active_disable() */
528 smp_rmb();
529
Zheng Yongjun974d8f82021-06-07 23:01:09 +0800530 /* Limit timeout to max: 2^6 * initial timeout */
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700531 multiplier = 1 << min(tfo_da_times - 1, 6);
Eric Dumazet6f20c8a2021-07-19 02:12:18 -0700532
533 /* Paired with the WRITE_ONCE() in tcp_fastopen_active_disable(). */
534 timeout = READ_ONCE(sock_net(sk)->ipv4.tfo_active_disable_stamp) +
535 multiplier * tfo_bh_timeout * HZ;
536 if (time_before(jiffies, timeout))
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700537 return true;
538
539 /* Mark check bit so we can check for successful active TFO
540 * condition and reset tfo_active_disable_times
541 */
542 tcp_sk(sk)->syn_fastopen_ch = 1;
543 return false;
544}
545
546/* Disable active TFO if FIN is the only packet in the ofo queue
547 * and no data is received.
548 * Also check if we can reset tfo_active_disable_times if data is
549 * received successfully on a marked active TFO sockets opened on
550 * a non-loopback interface
551 */
552void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
553{
554 struct tcp_sock *tp = tcp_sk(sk);
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700555 struct dst_entry *dst;
Eric Dumazet18a4c0e2017-10-05 22:21:21 -0700556 struct sk_buff *skb;
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700557
558 if (!tp->syn_fastopen)
559 return;
560
561 if (!tp->data_segs_in) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -0700562 skb = skb_rb_first(&tp->out_of_order_queue);
563 if (skb && !skb_rb_next(skb)) {
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700564 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
Wei Wang46c2fa32017-04-20 14:45:47 -0700565 tcp_fastopen_active_disable(sk);
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700566 return;
567 }
568 }
569 } else if (tp->syn_fastopen_ch &&
Haishuang Yan3733be12017-09-27 11:35:43 +0800570 atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700571 dst = sk_dst_get(sk);
572 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
Haishuang Yan3733be12017-09-27 11:35:43 +0800573 atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
Wei Wangcf1ef3f2017-04-20 14:45:46 -0700574 dst_release(dst);
575 }
576}
Yuchung Cheng72685862017-12-12 13:10:40 -0800577
578void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
579{
580 u32 timeouts = inet_csk(sk)->icsk_retransmits;
581 struct tcp_sock *tp = tcp_sk(sk);
582
583 /* Broken middle-boxes may black-hole Fast Open connection during or
584 * even after the handshake. Be extremely conservative and pause
585 * Fast Open globally after hitting the third consecutive timeout or
586 * exceeding the configured timeout limit.
587 */
588 if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
589 (timeouts == 2 || (timeouts < 2 && expired))) {
590 tcp_fastopen_active_disable(sk);
591 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
592 }
593}