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 request_sock
*req
)
177 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
181 req
->num_retrans
= 0;
182 req
->num_timeout
= 0;
185 child
= inet_csk(sk
)->icsk_af_ops
->syn_recv_sock(sk
, skb
, req
, NULL
,
190 spin_lock(&queue
->fastopenq
.lock
);
191 queue
->fastopenq
.qlen
++;
192 spin_unlock(&queue
->fastopenq
.lock
);
194 /* Initialize the child socket. Have to fix some values to take
195 * into account the child is a Fast Open socket and is created
196 * only out of the bits carried in the SYN packet.
200 tp
->fastopen_rsk
= req
;
201 tcp_rsk(req
)->tfo_listener
= true;
203 /* RFC1323: The window in SYN & SYN/ACK segments is never
204 * scaled. So correct it appropriately.
206 tp
->snd_wnd
= ntohs(tcp_hdr(skb
)->window
);
207 tp
->max_window
= tp
->snd_wnd
;
209 /* Activate the retrans timer so that SYNACK can be retransmitted.
210 * The request socket is not added to the ehash
211 * because it's been added to the accept queue directly.
213 inet_csk_reset_xmit_timer(child
, ICSK_TIME_RETRANS
,
214 TCP_TIMEOUT_INIT
, TCP_RTO_MAX
);
216 refcount_set(&req
->rsk_refcnt
, 2);
218 /* Now finish processing the fastopen child socket. */
219 inet_csk(child
)->icsk_af_ops
->rebuild_header(child
);
220 tcp_init_congestion_control(child
);
221 tcp_mtup_init(child
);
222 tcp_init_metrics(child
);
223 tcp_call_bpf(child
, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
);
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
)
282 struct tcp_fastopen_cookie valid_foc
= { .len
= -1 };
283 bool syn_data
= TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
+ 1;
286 if (foc
->len
== 0) /* Client requests a cookie */
287 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENCOOKIEREQD
);
289 if (!((sysctl_tcp_fastopen
& TFO_SERVER_ENABLE
) &&
290 (syn_data
|| foc
->len
>= 0) &&
291 tcp_fastopen_queue_check(sk
))) {
296 if (syn_data
&& (sysctl_tcp_fastopen
& TFO_SERVER_COOKIE_NOT_REQD
))
299 if (foc
->len
>= 0 && /* Client presents or requests a cookie */
300 tcp_fastopen_cookie_gen(req
, skb
, &valid_foc
) &&
301 foc
->len
== TCP_FASTOPEN_COOKIE_SIZE
&&
302 foc
->len
== valid_foc
.len
&&
303 !memcmp(foc
->val
, valid_foc
.val
, foc
->len
)) {
304 /* Cookie is valid. Create a (full) child socket to accept
305 * the data in SYN before returning a SYN-ACK to ack the
306 * data. If we fail to create the socket, fall back and
307 * ack the ISN only but includes the same cookie.
309 * Note: Data-less SYN with valid cookie is allowed to send
310 * data in SYN_RECV state.
313 child
= tcp_fastopen_create_child(sk
, skb
, req
);
316 NET_INC_STATS(sock_net(sk
),
317 LINUX_MIB_TCPFASTOPENPASSIVE
);
320 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENPASSIVEFAIL
);
321 } else if (foc
->len
> 0) /* Client presents an invalid cookie */
322 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENPASSIVEFAIL
);
324 valid_foc
.exp
= foc
->exp
;
329 bool tcp_fastopen_cookie_check(struct sock
*sk
, u16
*mss
,
330 struct tcp_fastopen_cookie
*cookie
)
332 unsigned long last_syn_loss
= 0;
335 tcp_fastopen_cache_get(sk
, mss
, cookie
, &syn_loss
, &last_syn_loss
);
337 /* Recurring FO SYN losses: no cookie or data in SYN */
339 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
344 /* Firewall blackhole issue check */
345 if (tcp_fastopen_active_should_disable(sk
)) {
350 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
) {
354 return cookie
->len
> 0;
357 /* This function checks if we want to defer sending SYN until the first
358 * write(). We defer under the following conditions:
359 * 1. fastopen_connect sockopt is set
360 * 2. we have a valid cookie
361 * Return value: return true if we want to defer until application writes data
362 * return false if we want to send out SYN immediately
364 bool tcp_fastopen_defer_connect(struct sock
*sk
, int *err
)
366 struct tcp_fastopen_cookie cookie
= { .len
= 0 };
367 struct tcp_sock
*tp
= tcp_sk(sk
);
370 if (tp
->fastopen_connect
&& !tp
->fastopen_req
) {
371 if (tcp_fastopen_cookie_check(sk
, &mss
, &cookie
)) {
372 inet_sk(sk
)->defer_connect
= 1;
376 /* Alloc fastopen_req in order for FO option to be included
379 tp
->fastopen_req
= kzalloc(sizeof(*tp
->fastopen_req
),
381 if (tp
->fastopen_req
)
382 tp
->fastopen_req
->cookie
= cookie
;
388 EXPORT_SYMBOL(tcp_fastopen_defer_connect
);
391 * The following code block is to deal with middle box issues with TFO:
392 * Middlebox firewall issues can potentially cause server's data being
393 * blackholed after a successful 3WHS using TFO.
394 * The proposed solution is to disable active TFO globally under the
395 * following circumstances:
396 * 1. client side TFO socket receives out of order FIN
397 * 2. client side TFO socket receives out of order RST
398 * We disable active side TFO globally for 1hr at first. Then if it
399 * happens again, we disable it for 2h, then 4h, 8h, ...
400 * And we reset the timeout back to 1hr when we see a successful active
401 * TFO connection with data exchanges.
405 unsigned int sysctl_tcp_fastopen_blackhole_timeout __read_mostly
= 60 * 60;
406 static atomic_t tfo_active_disable_times __read_mostly
= ATOMIC_INIT(0);
407 static unsigned long tfo_active_disable_stamp __read_mostly
;
409 /* Disable active TFO and record current jiffies and
410 * tfo_active_disable_times
412 void tcp_fastopen_active_disable(struct sock
*sk
)
414 atomic_inc(&tfo_active_disable_times
);
415 tfo_active_disable_stamp
= jiffies
;
416 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENBLACKHOLE
);
419 /* Reset tfo_active_disable_times to 0 */
420 void tcp_fastopen_active_timeout_reset(void)
422 atomic_set(&tfo_active_disable_times
, 0);
425 /* Calculate timeout for tfo active disable
426 * Return true if we are still in the active TFO disable period
427 * Return false if timeout already expired and we should use active TFO
429 bool tcp_fastopen_active_should_disable(struct sock
*sk
)
431 int tfo_da_times
= atomic_read(&tfo_active_disable_times
);
433 unsigned long timeout
;
438 /* Limit timout to max: 2^6 * initial timeout */
439 multiplier
= 1 << min(tfo_da_times
- 1, 6);
440 timeout
= multiplier
* sysctl_tcp_fastopen_blackhole_timeout
* HZ
;
441 if (time_before(jiffies
, tfo_active_disable_stamp
+ timeout
))
444 /* Mark check bit so we can check for successful active TFO
445 * condition and reset tfo_active_disable_times
447 tcp_sk(sk
)->syn_fastopen_ch
= 1;
451 /* Disable active TFO if FIN is the only packet in the ofo queue
452 * and no data is received.
453 * Also check if we can reset tfo_active_disable_times if data is
454 * received successfully on a marked active TFO sockets opened on
455 * a non-loopback interface
457 void tcp_fastopen_active_disable_ofo_check(struct sock
*sk
)
459 struct tcp_sock
*tp
= tcp_sk(sk
);
462 struct dst_entry
*dst
;
464 if (!tp
->syn_fastopen
)
467 if (!tp
->data_segs_in
) {
468 p
= rb_first(&tp
->out_of_order_queue
);
469 if (p
&& !rb_next(p
)) {
470 skb
= rb_entry(p
, struct sk_buff
, rbnode
);
471 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) {
472 tcp_fastopen_active_disable(sk
);
476 } else if (tp
->syn_fastopen_ch
&&
477 atomic_read(&tfo_active_disable_times
)) {
478 dst
= sk_dst_get(sk
);
479 if (!(dst
&& dst
->dev
&& (dst
->dev
->flags
& IFF_LOOPBACK
)))
480 tcp_fastopen_active_timeout_reset();