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 int sysctl_tcp_fastopen __read_mostly = TFO_CLIENT_ENABLE;
14 struct tcp_fastopen_context __rcu *tcp_fastopen_ctx;
16 static DEFINE_SPINLOCK(tcp_fastopen_ctx_lock);
18 void tcp_fastopen_init_key_once(bool publish)
20 static u8 key[TCP_FASTOPEN_KEY_LENGTH];
22 /* tcp_fastopen_reset_cipher publishes the new context
23 * atomically, so we allow this race happening here.
25 * All call sites of tcp_fastopen_cookie_gen also check
26 * for a valid cookie, so this is an acceptable risk.
28 if (net_get_random_once(key, sizeof(key)) && publish)
29 tcp_fastopen_reset_cipher(key, sizeof(key));
32 static void tcp_fastopen_ctx_free(struct rcu_head *head)
34 struct tcp_fastopen_context *ctx =
35 container_of(head, struct tcp_fastopen_context, rcu);
36 crypto_free_cipher(ctx->tfm);
40 int tcp_fastopen_reset_cipher(void *key, unsigned int len)
43 struct tcp_fastopen_context *ctx, *octx;
45 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
48 ctx->tfm = crypto_alloc_cipher("aes", 0, 0);
50 if (IS_ERR(ctx->tfm)) {
51 err = PTR_ERR(ctx->tfm);
53 pr_err("TCP: TFO aes cipher alloc error: %d\n", err);
56 err = crypto_cipher_setkey(ctx->tfm, key, len);
58 pr_err("TCP: TFO cipher key error: %d\n", err);
59 crypto_free_cipher(ctx->tfm);
62 memcpy(ctx->key, key, len);
64 spin_lock(&tcp_fastopen_ctx_lock);
66 octx = rcu_dereference_protected(tcp_fastopen_ctx,
67 lockdep_is_held(&tcp_fastopen_ctx_lock));
68 rcu_assign_pointer(tcp_fastopen_ctx, ctx);
69 spin_unlock(&tcp_fastopen_ctx_lock);
72 call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
76 static bool __tcp_fastopen_cookie_gen(const void *path,
77 struct tcp_fastopen_cookie *foc)
79 struct tcp_fastopen_context *ctx;
83 ctx = rcu_dereference(tcp_fastopen_ctx);
85 crypto_cipher_encrypt_one(ctx->tfm, foc->val, path);
86 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
93 /* Generate the fastopen cookie by doing aes128 encryption on both
94 * the source and destination addresses. Pad 0s for IPv4 or IPv4-mapped-IPv6
95 * addresses. For the longer IPv6 addresses use CBC-MAC.
97 * XXX (TFO) - refactor when TCP_FASTOPEN_COOKIE_SIZE != AES_BLOCK_SIZE.
99 static bool tcp_fastopen_cookie_gen(struct request_sock *req,
101 struct tcp_fastopen_cookie *foc)
103 if (req->rsk_ops->family == AF_INET) {
104 const struct iphdr *iph = ip_hdr(syn);
106 __be32 path[4] = { iph->saddr, iph->daddr, 0, 0 };
107 return __tcp_fastopen_cookie_gen(path, foc);
110 #if IS_ENABLED(CONFIG_IPV6)
111 if (req->rsk_ops->family == AF_INET6) {
112 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
113 struct tcp_fastopen_cookie tmp;
115 if (__tcp_fastopen_cookie_gen(&ip6h->saddr, &tmp)) {
116 struct in6_addr *buf = &tmp.addr;
119 for (i = 0; i < 4; i++)
120 buf->s6_addr32[i] ^= ip6h->daddr.s6_addr32[i];
121 return __tcp_fastopen_cookie_gen(buf, foc);
129 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
130 * queue this additional data / FIN.
132 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
134 struct tcp_sock *tp = tcp_sk(sk);
136 if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
139 skb = skb_clone(skb, GFP_ATOMIC);
144 /* segs_in has been initialized to 1 in tcp_create_openreq_child().
145 * Hence, reset segs_in to 0 before calling tcp_segs_in()
146 * to avoid double counting. Also, tcp_segs_in() expects
147 * skb->len to include the tcp_hdrlen. Hence, it should
148 * be called before __skb_pull().
151 tcp_segs_in(tp, skb);
152 __skb_pull(skb, tcp_hdrlen(skb));
153 sk_forced_mem_schedule(sk, skb->truesize);
154 skb_set_owner_r(skb, sk);
156 TCP_SKB_CB(skb)->seq++;
157 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
159 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
160 __skb_queue_tail(&sk->sk_receive_queue, skb);
161 tp->syn_data_acked = 1;
163 /* u64_stats_update_begin(&tp->syncp) not needed here,
164 * as we certainly are not changing upper 32bit value (0)
166 tp->bytes_received = skb->len;
168 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
172 static struct sock *tcp_fastopen_create_child(struct sock *sk,
174 struct dst_entry *dst,
175 struct request_sock *req)
178 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
182 req->num_retrans = 0;
183 req->num_timeout = 0;
186 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
191 spin_lock(&queue->fastopenq.lock);
192 queue->fastopenq.qlen++;
193 spin_unlock(&queue->fastopenq.lock);
195 /* Initialize the child socket. Have to fix some values to take
196 * into account the child is a Fast Open socket and is created
197 * only out of the bits carried in the SYN packet.
201 tp->fastopen_rsk = req;
202 tcp_rsk(req)->tfo_listener = true;
204 /* RFC1323: The window in SYN & SYN/ACK segments is never
205 * scaled. So correct it appropriately.
207 tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
208 tp->max_window = tp->snd_wnd;
210 /* Activate the retrans timer so that SYNACK can be retransmitted.
211 * The request socket is not added to the ehash
212 * because it's been added to the accept queue directly.
214 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
215 TCP_TIMEOUT_INIT, TCP_RTO_MAX);
217 atomic_set(&req->rsk_refcnt, 2);
219 /* Now finish processing the fastopen child socket. */
220 inet_csk(child)->icsk_af_ops->rebuild_header(child);
221 tcp_init_congestion_control(child);
222 tcp_mtup_init(child);
223 tcp_init_metrics(child);
224 tcp_init_buffer_space(child);
226 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
228 tcp_fastopen_add_skb(child, skb);
230 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
231 tp->rcv_wup = tp->rcv_nxt;
232 /* tcp_conn_request() is sending the SYNACK,
233 * and queues the child into listener accept queue.
238 static bool tcp_fastopen_queue_check(struct sock *sk)
240 struct fastopen_queue *fastopenq;
242 /* Make sure the listener has enabled fastopen, and we don't
243 * exceed the max # of pending TFO requests allowed before trying
244 * to validating the cookie in order to avoid burning CPU cycles
247 * XXX (TFO) - The implication of checking the max_qlen before
248 * processing a cookie request is that clients can't differentiate
249 * between qlen overflow causing Fast Open to be disabled
250 * temporarily vs a server not supporting Fast Open at all.
252 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
253 if (fastopenq->max_qlen == 0)
256 if (fastopenq->qlen >= fastopenq->max_qlen) {
257 struct request_sock *req1;
258 spin_lock(&fastopenq->lock);
259 req1 = fastopenq->rskq_rst_head;
260 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
261 __NET_INC_STATS(sock_net(sk),
262 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
263 spin_unlock(&fastopenq->lock);
266 fastopenq->rskq_rst_head = req1->dl_next;
268 spin_unlock(&fastopenq->lock);
274 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
275 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
276 * cookie request (foc->len == 0).
278 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
279 struct request_sock *req,
280 struct tcp_fastopen_cookie *foc,
281 struct dst_entry *dst)
283 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
284 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
287 if (foc->len == 0) /* Client requests a cookie */
288 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
290 if (!((sysctl_tcp_fastopen & TFO_SERVER_ENABLE) &&
291 (syn_data || foc->len >= 0) &&
292 tcp_fastopen_queue_check(sk))) {
297 if (syn_data && (sysctl_tcp_fastopen & TFO_SERVER_COOKIE_NOT_REQD))
300 if (foc->len >= 0 && /* Client presents or requests a cookie */
301 tcp_fastopen_cookie_gen(req, skb, &valid_foc) &&
302 foc->len == TCP_FASTOPEN_COOKIE_SIZE &&
303 foc->len == valid_foc.len &&
304 !memcmp(foc->val, valid_foc.val, foc->len)) {
305 /* Cookie is valid. Create a (full) child socket to accept
306 * the data in SYN before returning a SYN-ACK to ack the
307 * data. If we fail to create the socket, fall back and
308 * ack the ISN only but includes the same cookie.
310 * Note: Data-less SYN with valid cookie is allowed to send
311 * data in SYN_RECV state.
314 child = tcp_fastopen_create_child(sk, skb, dst, req);
317 NET_INC_STATS(sock_net(sk),
318 LINUX_MIB_TCPFASTOPENPASSIVE);
321 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
322 } else if (foc->len > 0) /* Client presents an invalid cookie */
323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
325 valid_foc.exp = foc->exp;
330 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
331 struct tcp_fastopen_cookie *cookie)
333 unsigned long last_syn_loss = 0;
336 tcp_fastopen_cache_get(sk, mss, cookie, &syn_loss, &last_syn_loss);
338 /* Recurring FO SYN losses: no cookie or data in SYN */
340 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
345 /* Firewall blackhole issue check */
346 if (tcp_fastopen_active_should_disable(sk)) {
351 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) {
355 return cookie->len > 0;
358 /* This function checks if we want to defer sending SYN until the first
359 * write(). We defer under the following conditions:
360 * 1. fastopen_connect sockopt is set
361 * 2. we have a valid cookie
362 * Return value: return true if we want to defer until application writes data
363 * return false if we want to send out SYN immediately
365 bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
367 struct tcp_fastopen_cookie cookie = { .len = 0 };
368 struct tcp_sock *tp = tcp_sk(sk);
371 if (tp->fastopen_connect && !tp->fastopen_req) {
372 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
373 inet_sk(sk)->defer_connect = 1;
377 /* Alloc fastopen_req in order for FO option to be included
380 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
382 if (tp->fastopen_req)
383 tp->fastopen_req->cookie = cookie;
389 EXPORT_SYMBOL(tcp_fastopen_defer_connect);
392 * The following code block is to deal with middle box issues with TFO:
393 * Middlebox firewall issues can potentially cause server's data being
394 * blackholed after a successful 3WHS using TFO.
395 * The proposed solution is to disable active TFO globally under the
396 * following circumstances:
397 * 1. client side TFO socket receives out of order FIN
398 * 2. client side TFO socket receives out of order RST
399 * We disable active side TFO globally for 1hr at first. Then if it
400 * happens again, we disable it for 2h, then 4h, 8h, ...
401 * And we reset the timeout back to 1hr when we see a successful active
402 * TFO connection with data exchanges.
406 unsigned int sysctl_tcp_fastopen_blackhole_timeout __read_mostly = 60 * 60;
407 static atomic_t tfo_active_disable_times __read_mostly = ATOMIC_INIT(0);
408 static unsigned long tfo_active_disable_stamp __read_mostly;
410 /* Disable active TFO and record current jiffies and
411 * tfo_active_disable_times
413 void tcp_fastopen_active_disable(struct sock *sk)
415 atomic_inc(&tfo_active_disable_times);
416 tfo_active_disable_stamp = jiffies;
417 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENBLACKHOLE);
420 /* Reset tfo_active_disable_times to 0 */
421 void tcp_fastopen_active_timeout_reset(void)
423 atomic_set(&tfo_active_disable_times, 0);
426 /* Calculate timeout for tfo active disable
427 * Return true if we are still in the active TFO disable period
428 * Return false if timeout already expired and we should use active TFO
430 bool tcp_fastopen_active_should_disable(struct sock *sk)
432 int tfo_da_times = atomic_read(&tfo_active_disable_times);
434 unsigned long timeout;
439 /* Limit timout to max: 2^6 * initial timeout */
440 multiplier = 1 << min(tfo_da_times - 1, 6);
441 timeout = multiplier * sysctl_tcp_fastopen_blackhole_timeout * HZ;
442 if (time_before(jiffies, tfo_active_disable_stamp + timeout))
445 /* Mark check bit so we can check for successful active TFO
446 * condition and reset tfo_active_disable_times
448 tcp_sk(sk)->syn_fastopen_ch = 1;
452 /* Disable active TFO if FIN is the only packet in the ofo queue
453 * and no data is received.
454 * Also check if we can reset tfo_active_disable_times if data is
455 * received successfully on a marked active TFO sockets opened on
456 * a non-loopback interface
458 void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
460 struct tcp_sock *tp = tcp_sk(sk);
463 struct dst_entry *dst;
465 if (!tp->syn_fastopen)
468 if (!tp->data_segs_in) {
469 p = rb_first(&tp->out_of_order_queue);
470 if (p && !rb_next(p)) {
471 skb = rb_entry(p, struct sk_buff, rbnode);
472 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
473 tcp_fastopen_active_disable(sk);
477 } else if (tp->syn_fastopen_ch &&
478 atomic_read(&tfo_active_disable_times)) {
479 dst = sk_dst_get(sk);
480 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
481 tcp_fastopen_active_timeout_reset();