1 #include <linux/crypto.h>
3 #include <linux/init.h>
4 #include <linux/kernel.h>
5 #include <linux/list.h>
7 #include <linux/rcupdate.h>
8 #include <linux/rculist.h>
9 #include <net/inetpeer.h>
12 void tcp_fastopen_init_key_once(struct net *net)
14 u8 key[TCP_FASTOPEN_KEY_LENGTH];
15 struct tcp_fastopen_context *ctxt;
18 ctxt = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
25 /* tcp_fastopen_reset_cipher publishes the new context
26 * atomically, so we allow this race happening here.
28 * All call sites of tcp_fastopen_cookie_gen also check
29 * for a valid cookie, so this is an acceptable risk.
31 get_random_bytes(key, sizeof(key));
32 tcp_fastopen_reset_cipher(net, key, sizeof(key));
35 static void tcp_fastopen_ctx_free(struct rcu_head *head)
37 struct tcp_fastopen_context *ctx =
38 container_of(head, struct tcp_fastopen_context, rcu);
39 crypto_free_cipher(ctx->tfm);
43 void tcp_fastopen_ctx_destroy(struct net *net)
45 struct tcp_fastopen_context *ctxt;
47 spin_lock(&net->ipv4.tcp_fastopen_ctx_lock);
49 ctxt = rcu_dereference_protected(net->ipv4.tcp_fastopen_ctx,
50 lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
51 rcu_assign_pointer(net->ipv4.tcp_fastopen_ctx, NULL);
52 spin_unlock(&net->ipv4.tcp_fastopen_ctx_lock);
55 call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free);
58 int tcp_fastopen_reset_cipher(struct net *net, void *key, unsigned int len)
61 struct tcp_fastopen_context *ctx, *octx;
63 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
66 ctx->tfm = crypto_alloc_cipher("aes", 0, 0);
68 if (IS_ERR(ctx->tfm)) {
69 err = PTR_ERR(ctx->tfm);
71 pr_err("TCP: TFO aes cipher alloc error: %d\n", err);
74 err = crypto_cipher_setkey(ctx->tfm, key, len);
76 pr_err("TCP: TFO cipher key error: %d\n", err);
77 crypto_free_cipher(ctx->tfm);
80 memcpy(ctx->key, key, len);
82 spin_lock(&net->ipv4.tcp_fastopen_ctx_lock);
84 octx = rcu_dereference_protected(net->ipv4.tcp_fastopen_ctx,
85 lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
86 rcu_assign_pointer(net->ipv4.tcp_fastopen_ctx, ctx);
87 spin_unlock(&net->ipv4.tcp_fastopen_ctx_lock);
90 call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
94 static bool __tcp_fastopen_cookie_gen(struct net *net,
96 struct tcp_fastopen_cookie *foc)
98 struct tcp_fastopen_context *ctx;
102 ctx = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
104 crypto_cipher_encrypt_one(ctx->tfm, foc->val, path);
105 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
112 /* Generate the fastopen cookie by doing aes128 encryption on both
113 * the source and destination addresses. Pad 0s for IPv4 or IPv4-mapped-IPv6
114 * addresses. For the longer IPv6 addresses use CBC-MAC.
116 * XXX (TFO) - refactor when TCP_FASTOPEN_COOKIE_SIZE != AES_BLOCK_SIZE.
118 static bool tcp_fastopen_cookie_gen(struct net *net,
119 struct request_sock *req,
121 struct tcp_fastopen_cookie *foc)
123 if (req->rsk_ops->family == AF_INET) {
124 const struct iphdr *iph = ip_hdr(syn);
126 __be32 path[4] = { iph->saddr, iph->daddr, 0, 0 };
127 return __tcp_fastopen_cookie_gen(net, path, foc);
130 #if IS_ENABLED(CONFIG_IPV6)
131 if (req->rsk_ops->family == AF_INET6) {
132 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
133 struct tcp_fastopen_cookie tmp;
135 if (__tcp_fastopen_cookie_gen(net, &ip6h->saddr, &tmp)) {
136 struct in6_addr *buf = &tmp.addr;
139 for (i = 0; i < 4; i++)
140 buf->s6_addr32[i] ^= ip6h->daddr.s6_addr32[i];
141 return __tcp_fastopen_cookie_gen(net, buf, foc);
149 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
150 * queue this additional data / FIN.
152 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
154 struct tcp_sock *tp = tcp_sk(sk);
156 if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
159 skb = skb_clone(skb, GFP_ATOMIC);
164 /* segs_in has been initialized to 1 in tcp_create_openreq_child().
165 * Hence, reset segs_in to 0 before calling tcp_segs_in()
166 * to avoid double counting. Also, tcp_segs_in() expects
167 * skb->len to include the tcp_hdrlen. Hence, it should
168 * be called before __skb_pull().
171 tcp_segs_in(tp, skb);
172 __skb_pull(skb, tcp_hdrlen(skb));
173 sk_forced_mem_schedule(sk, skb->truesize);
174 skb_set_owner_r(skb, sk);
176 TCP_SKB_CB(skb)->seq++;
177 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
179 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
180 __skb_queue_tail(&sk->sk_receive_queue, skb);
181 tp->syn_data_acked = 1;
183 /* u64_stats_update_begin(&tp->syncp) not needed here,
184 * as we certainly are not changing upper 32bit value (0)
186 tp->bytes_received = skb->len;
188 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
192 static struct sock *tcp_fastopen_create_child(struct sock *sk,
194 struct request_sock *req)
197 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
201 req->num_retrans = 0;
202 req->num_timeout = 0;
205 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
210 spin_lock(&queue->fastopenq.lock);
211 queue->fastopenq.qlen++;
212 spin_unlock(&queue->fastopenq.lock);
214 /* Initialize the child socket. Have to fix some values to take
215 * into account the child is a Fast Open socket and is created
216 * only out of the bits carried in the SYN packet.
220 tp->fastopen_rsk = req;
221 tcp_rsk(req)->tfo_listener = true;
223 /* RFC1323: The window in SYN & SYN/ACK segments is never
224 * scaled. So correct it appropriately.
226 tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
227 tp->max_window = tp->snd_wnd;
229 /* Activate the retrans timer so that SYNACK can be retransmitted.
230 * The request socket is not added to the ehash
231 * because it's been added to the accept queue directly.
233 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
234 TCP_TIMEOUT_INIT, TCP_RTO_MAX);
236 refcount_set(&req->rsk_refcnt, 2);
238 /* Now finish processing the fastopen child socket. */
239 inet_csk(child)->icsk_af_ops->rebuild_header(child);
240 tcp_init_congestion_control(child);
241 tcp_mtup_init(child);
242 tcp_init_metrics(child);
243 tcp_call_bpf(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
244 tcp_init_buffer_space(child);
246 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
248 tcp_fastopen_add_skb(child, skb);
250 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
251 tp->rcv_wup = tp->rcv_nxt;
252 /* tcp_conn_request() is sending the SYNACK,
253 * and queues the child into listener accept queue.
258 static bool tcp_fastopen_queue_check(struct sock *sk)
260 struct fastopen_queue *fastopenq;
262 /* Make sure the listener has enabled fastopen, and we don't
263 * exceed the max # of pending TFO requests allowed before trying
264 * to validating the cookie in order to avoid burning CPU cycles
267 * XXX (TFO) - The implication of checking the max_qlen before
268 * processing a cookie request is that clients can't differentiate
269 * between qlen overflow causing Fast Open to be disabled
270 * temporarily vs a server not supporting Fast Open at all.
272 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
273 if (fastopenq->max_qlen == 0)
276 if (fastopenq->qlen >= fastopenq->max_qlen) {
277 struct request_sock *req1;
278 spin_lock(&fastopenq->lock);
279 req1 = fastopenq->rskq_rst_head;
280 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
281 __NET_INC_STATS(sock_net(sk),
282 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
283 spin_unlock(&fastopenq->lock);
286 fastopenq->rskq_rst_head = req1->dl_next;
288 spin_unlock(&fastopenq->lock);
294 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
295 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
296 * cookie request (foc->len == 0).
298 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
299 struct request_sock *req,
300 struct tcp_fastopen_cookie *foc)
302 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
303 int tcp_fastopen = sock_net(sk)->ipv4.sysctl_tcp_fastopen;
304 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
307 if (foc->len == 0) /* Client requests a cookie */
308 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
310 if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
311 (syn_data || foc->len >= 0) &&
312 tcp_fastopen_queue_check(sk))) {
317 if (syn_data && (tcp_fastopen & TFO_SERVER_COOKIE_NOT_REQD))
320 if (foc->len >= 0 && /* Client presents or requests a cookie */
321 tcp_fastopen_cookie_gen(sock_net(sk), req, skb, &valid_foc) &&
322 foc->len == TCP_FASTOPEN_COOKIE_SIZE &&
323 foc->len == valid_foc.len &&
324 !memcmp(foc->val, valid_foc.val, foc->len)) {
325 /* Cookie is valid. Create a (full) child socket to accept
326 * the data in SYN before returning a SYN-ACK to ack the
327 * data. If we fail to create the socket, fall back and
328 * ack the ISN only but includes the same cookie.
330 * Note: Data-less SYN with valid cookie is allowed to send
331 * data in SYN_RECV state.
334 child = tcp_fastopen_create_child(sk, skb, req);
337 NET_INC_STATS(sock_net(sk),
338 LINUX_MIB_TCPFASTOPENPASSIVE);
341 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
342 } else if (foc->len > 0) /* Client presents an invalid cookie */
343 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
345 valid_foc.exp = foc->exp;
350 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
351 struct tcp_fastopen_cookie *cookie)
353 unsigned long last_syn_loss = 0;
356 tcp_fastopen_cache_get(sk, mss, cookie, &syn_loss, &last_syn_loss);
358 /* Recurring FO SYN losses: no cookie or data in SYN */
360 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
365 /* Firewall blackhole issue check */
366 if (tcp_fastopen_active_should_disable(sk)) {
371 if (sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) {
375 return cookie->len > 0;
378 /* This function checks if we want to defer sending SYN until the first
379 * write(). We defer under the following conditions:
380 * 1. fastopen_connect sockopt is set
381 * 2. we have a valid cookie
382 * Return value: return true if we want to defer until application writes data
383 * return false if we want to send out SYN immediately
385 bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
387 struct tcp_fastopen_cookie cookie = { .len = 0 };
388 struct tcp_sock *tp = tcp_sk(sk);
391 if (tp->fastopen_connect && !tp->fastopen_req) {
392 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
393 inet_sk(sk)->defer_connect = 1;
397 /* Alloc fastopen_req in order for FO option to be included
400 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
402 if (tp->fastopen_req)
403 tp->fastopen_req->cookie = cookie;
409 EXPORT_SYMBOL(tcp_fastopen_defer_connect);
412 * The following code block is to deal with middle box issues with TFO:
413 * Middlebox firewall issues can potentially cause server's data being
414 * blackholed after a successful 3WHS using TFO.
415 * The proposed solution is to disable active TFO globally under the
416 * following circumstances:
417 * 1. client side TFO socket receives out of order FIN
418 * 2. client side TFO socket receives out of order RST
419 * We disable active side TFO globally for 1hr at first. Then if it
420 * happens again, we disable it for 2h, then 4h, 8h, ...
421 * And we reset the timeout back to 1hr when we see a successful active
422 * TFO connection with data exchanges.
425 /* Disable active TFO and record current jiffies and
426 * tfo_active_disable_times
428 void tcp_fastopen_active_disable(struct sock *sk)
430 struct net *net = sock_net(sk);
432 atomic_inc(&net->ipv4.tfo_active_disable_times);
433 net->ipv4.tfo_active_disable_stamp = jiffies;
434 NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
437 /* Calculate timeout for tfo active disable
438 * Return true if we are still in the active TFO disable period
439 * Return false if timeout already expired and we should use active TFO
441 bool tcp_fastopen_active_should_disable(struct sock *sk)
443 unsigned int tfo_bh_timeout = sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout;
444 int tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
445 unsigned long timeout;
451 /* Limit timout to max: 2^6 * initial timeout */
452 multiplier = 1 << min(tfo_da_times - 1, 6);
453 timeout = multiplier * tfo_bh_timeout * HZ;
454 if (time_before(jiffies, sock_net(sk)->ipv4.tfo_active_disable_stamp + timeout))
457 /* Mark check bit so we can check for successful active TFO
458 * condition and reset tfo_active_disable_times
460 tcp_sk(sk)->syn_fastopen_ch = 1;
464 /* Disable active TFO if FIN is the only packet in the ofo queue
465 * and no data is received.
466 * Also check if we can reset tfo_active_disable_times if data is
467 * received successfully on a marked active TFO sockets opened on
468 * a non-loopback interface
470 void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
472 struct tcp_sock *tp = tcp_sk(sk);
475 struct dst_entry *dst;
477 if (!tp->syn_fastopen)
480 if (!tp->data_segs_in) {
481 p = rb_first(&tp->out_of_order_queue);
482 if (p && !rb_next(p)) {
483 skb = rb_entry(p, struct sk_buff, rbnode);
484 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
485 tcp_fastopen_active_disable(sk);
489 } else if (tp->syn_fastopen_ch &&
490 atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
491 dst = sk_dst_get(sk);
492 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
493 atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);