Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
2100c8d2 | 2 | #include <linux/kernel.h> |
10467163 JC |
3 | #include <linux/tcp.h> |
4 | #include <linux/rcupdate.h> | |
10467163 | 5 | #include <net/tcp.h> |
2100c8d2 | 6 | |
43713848 | 7 | void tcp_fastopen_init_key_once(struct net *net) |
222e83d2 | 8 | { |
43713848 HY |
9 | 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(); | |
222e83d2 HFS |
19 | |
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 | */ | |
43713848 | 26 | get_random_bytes(key, sizeof(key)); |
438ac880 | 27 | tcp_fastopen_reset_cipher(net, NULL, key, NULL); |
222e83d2 HFS |
28 | } |
29 | ||
10467163 JC |
30 | static 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); | |
9092a76d | 34 | |
453431a5 | 35 | kfree_sensitive(ctx); |
10467163 JC |
36 | } |
37 | ||
1fba70e5 YC |
38 | void 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 | ||
43713848 HY |
48 | void tcp_fastopen_ctx_destroy(struct net *net) |
49 | { | |
50 | struct tcp_fastopen_context *ctxt; | |
51 | ||
e93abb84 | 52 | ctxt = xchg((__force struct tcp_fastopen_context **)&net->ipv4.tcp_fastopen_ctx, NULL); |
43713848 HY |
53 | |
54 | if (ctxt) | |
55 | call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free); | |
56 | } | |
57 | ||
1fba70e5 | 58 | int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk, |
438ac880 | 59 | void *primary_key, void *backup_key) |
10467163 | 60 | { |
10467163 | 61 | struct tcp_fastopen_context *ctx, *octx; |
1fba70e5 | 62 | struct fastopen_queue *q; |
9092a76d | 63 | int err = 0; |
10467163 | 64 | |
c681edae AB |
65 | ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); |
66 | if (!ctx) { | |
67 | err = -ENOMEM; | |
9092a76d | 68 | goto out; |
10467163 | 69 | } |
c681edae | 70 | |
438ac880 AB |
71 | ctx->key[0].key[0] = get_unaligned_le64(primary_key); |
72 | ctx->key[0].key[1] = get_unaligned_le64(primary_key + 8); | |
c681edae | 73 | if (backup_key) { |
438ac880 AB |
74 | ctx->key[1].key[0] = get_unaligned_le64(backup_key); |
75 | ctx->key[1].key[1] = get_unaligned_le64(backup_key + 8); | |
c681edae AB |
76 | ctx->num = 2; |
77 | } else { | |
78 | ctx->num = 1; | |
79 | } | |
80 | ||
1fba70e5 YC |
81 | if (sk) { |
82 | q = &inet_csk(sk)->icsk_accept_queue.fastopenq; | |
e93abb84 | 83 | octx = xchg((__force struct tcp_fastopen_context **)&q->ctx, ctx); |
1fba70e5 | 84 | } else { |
e93abb84 | 85 | octx = xchg((__force struct tcp_fastopen_context **)&net->ipv4.tcp_fastopen_ctx, ctx); |
1fba70e5 | 86 | } |
10467163 JC |
87 | |
88 | if (octx) | |
89 | call_rcu(&octx->rcu, tcp_fastopen_ctx_free); | |
9092a76d | 90 | out: |
10467163 JC |
91 | return err; |
92 | } | |
93 | ||
f19008e6 JB |
94 | int 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 | ||
483642e5 CP |
117 | static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req, |
118 | struct sk_buff *syn, | |
438ac880 | 119 | const siphash_key_t *key, |
483642e5 | 120 | struct tcp_fastopen_cookie *foc) |
10467163 | 121 | { |
c681edae AB |
122 | BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64)); |
123 | ||
483642e5 CP |
124 | if (req->rsk_ops->family == AF_INET) { |
125 | const struct iphdr *iph = ip_hdr(syn); | |
1fba70e5 | 126 | |
438ac880 AB |
127 | foc->val[0] = cpu_to_le64(siphash(&iph->saddr, |
128 | sizeof(iph->saddr) + | |
129 | sizeof(iph->daddr), | |
130 | key)); | |
483642e5 CP |
131 | foc->len = TCP_FASTOPEN_COOKIE_SIZE; |
132 | return true; | |
133 | } | |
483642e5 CP |
134 | #if IS_ENABLED(CONFIG_IPV6) |
135 | if (req->rsk_ops->family == AF_INET6) { | |
136 | const struct ipv6hdr *ip6h = ipv6_hdr(syn); | |
c681edae | 137 | |
438ac880 AB |
138 | foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr, |
139 | sizeof(ip6h->saddr) + | |
140 | sizeof(ip6h->daddr), | |
141 | key)); | |
10467163 | 142 | foc->len = TCP_FASTOPEN_COOKIE_SIZE; |
483642e5 | 143 | return true; |
10467163 | 144 | } |
483642e5 CP |
145 | #endif |
146 | return false; | |
3a19ce0e DL |
147 | } |
148 | ||
c681edae AB |
149 | /* Generate the fastopen cookie by applying SipHash to both the source and |
150 | * destination addresses. | |
3a19ce0e | 151 | */ |
9092a76d | 152 | static void tcp_fastopen_cookie_gen(struct sock *sk, |
43713848 | 153 | struct request_sock *req, |
3a19ce0e DL |
154 | struct sk_buff *syn, |
155 | struct tcp_fastopen_cookie *foc) | |
156 | { | |
483642e5 | 157 | struct tcp_fastopen_context *ctx; |
3a19ce0e | 158 | |
483642e5 | 159 | rcu_read_lock(); |
9092a76d | 160 | ctx = tcp_fastopen_get_ctx(sk); |
483642e5 | 161 | if (ctx) |
438ac880 | 162 | __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc); |
483642e5 | 163 | rcu_read_unlock(); |
10467163 | 164 | } |
5b7ed089 | 165 | |
61d2bcae ED |
166 | /* If an incoming SYN or SYNACK frame contains a payload and/or FIN, |
167 | * queue this additional data / FIN. | |
168 | */ | |
169 | void 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); | |
a44d6eac MKL |
181 | /* 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); | |
61d2bcae | 189 | __skb_pull(skb, tcp_hdrlen(skb)); |
76061f63 | 190 | sk_forced_mem_schedule(sk, skb->truesize); |
61d2bcae ED |
191 | skb_set_owner_r(skb, sk); |
192 | ||
9d691539 ED |
193 | TCP_SKB_CB(skb)->seq++; |
194 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN; | |
195 | ||
61d2bcae ED |
196 | 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; | |
e3e17b77 ED |
204 | |
205 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) | |
206 | tcp_fin(sk); | |
61d2bcae ED |
207 | } |
208 | ||
9092a76d JB |
209 | /* returns 0 - no key match, 1 for primary, 2 for backup */ |
210 | static 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++) { | |
438ac880 | 226 | __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[i], foc); |
9092a76d JB |
227 | if (tcp_fastopen_cookie_match(foc, orig)) { |
228 | ret = i + 1; | |
229 | goto out; | |
230 | } | |
231 | foc = &search_foc; | |
232 | } | |
233 | out: | |
234 | rcu_read_unlock(); | |
235 | return ret; | |
236 | } | |
237 | ||
7c85af88 ED |
238 | static struct sock *tcp_fastopen_create_child(struct sock *sk, |
239 | struct sk_buff *skb, | |
7c85af88 | 240 | struct request_sock *req) |
5b7ed089 | 241 | { |
17846376 | 242 | struct tcp_sock *tp; |
5b7ed089 | 243 | struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; |
5b7ed089 | 244 | struct sock *child; |
5e0724d0 | 245 | bool own_req; |
5b7ed089 | 246 | |
5e0724d0 ED |
247 | child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, |
248 | NULL, &own_req); | |
51456b29 | 249 | if (!child) |
7c85af88 | 250 | return NULL; |
5b7ed089 | 251 | |
0536fcc0 ED |
252 | spin_lock(&queue->fastopenq.lock); |
253 | queue->fastopenq.qlen++; | |
254 | spin_unlock(&queue->fastopenq.lock); | |
5b7ed089 YC |
255 | |
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 | ||
d983ea6f | 262 | rcu_assign_pointer(tp->fastopen_rsk, req); |
9439ce00 | 263 | tcp_rsk(req)->tfo_listener = true; |
5b7ed089 YC |
264 | |
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); | |
0dbd7ff3 | 269 | tp->max_window = tp->snd_wnd; |
5b7ed089 YC |
270 | |
271 | /* Activate the retrans timer so that SYNACK can be retransmitted. | |
ca6fb065 | 272 | * The request socket is not added to the ehash |
5b7ed089 YC |
273 | * because it's been added to the accept queue directly. |
274 | */ | |
8ea731d4 | 275 | req->timeout = tcp_timeout_init(child); |
5b7ed089 | 276 | inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS, |
8ea731d4 | 277 | req->timeout, TCP_RTO_MAX); |
5b7ed089 | 278 | |
41c6d650 | 279 | refcount_set(&req->rsk_refcnt, 2); |
5b7ed089 YC |
280 | |
281 | /* Now finish processing the fastopen child socket. */ | |
72be0fe6 | 282 | tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, skb); |
5b7ed089 | 283 | |
61d2bcae ED |
284 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; |
285 | ||
286 | tcp_fastopen_add_skb(child, skb); | |
287 | ||
288 | tcp_rsk(req)->rcv_nxt = tp->rcv_nxt; | |
28b346cb | 289 | tp->rcv_wup = tp->rcv_nxt; |
7656d842 ED |
290 | /* tcp_conn_request() is sending the SYNACK, |
291 | * and queues the child into listener accept queue. | |
7c85af88 | 292 | */ |
7c85af88 | 293 | return child; |
5b7ed089 | 294 | } |
5b7ed089 YC |
295 | |
296 | static 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 | */ | |
0536fcc0 ED |
310 | fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq; |
311 | if (fastopenq->max_qlen == 0) | |
5b7ed089 YC |
312 | 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; | |
fa76ce73 | 318 | if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) { |
02a1d6e7 ED |
319 | __NET_INC_STATS(sock_net(sk), |
320 | LINUX_MIB_TCPFASTOPENLISTENOVERFLOW); | |
c10d9310 | 321 | spin_unlock(&fastopenq->lock); |
5b7ed089 YC |
322 | return false; |
323 | } | |
324 | fastopenq->rskq_rst_head = req1->dl_next; | |
325 | fastopenq->qlen--; | |
326 | spin_unlock(&fastopenq->lock); | |
13854e5a | 327 | reqsk_put(req1); |
5b7ed089 YC |
328 | } |
329 | return true; | |
330 | } | |
331 | ||
71c02379 CP |
332 | static bool tcp_fastopen_no_cookie(const struct sock *sk, |
333 | const struct dst_entry *dst, | |
334 | int flag) | |
335 | { | |
5a542133 | 336 | return (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) & flag) || |
71c02379 CP |
337 | tcp_sk(sk)->fastopen_no_cookie || |
338 | (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE)); | |
339 | } | |
340 | ||
89278c9d YC |
341 | /* 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 | */ | |
7c85af88 ED |
345 | struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb, |
346 | struct request_sock *req, | |
71c02379 CP |
347 | struct tcp_fastopen_cookie *foc, |
348 | const struct dst_entry *dst) | |
5b7ed089 | 349 | { |
89278c9d | 350 | bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1; |
5a542133 | 351 | int tcp_fastopen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen); |
e1cfcbe8 | 352 | struct tcp_fastopen_cookie valid_foc = { .len = -1 }; |
7c85af88 | 353 | struct sock *child; |
9092a76d | 354 | int ret = 0; |
5b7ed089 | 355 | |
531c94a9 | 356 | if (foc->len == 0) /* Client requests a cookie */ |
c10d9310 | 357 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD); |
531c94a9 | 358 | |
e1cfcbe8 | 359 | if (!((tcp_fastopen & TFO_SERVER_ENABLE) && |
89278c9d YC |
360 | (syn_data || foc->len >= 0) && |
361 | tcp_fastopen_queue_check(sk))) { | |
362 | foc->len = -1; | |
7c85af88 | 363 | return NULL; |
5b7ed089 YC |
364 | } |
365 | ||
e3faa49b | 366 | if (tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD)) |
89278c9d YC |
367 | goto fastopen; |
368 | ||
9092a76d JB |
369 | 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) { | |
376 | NET_INC_STATS(sock_net(sk), | |
377 | 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 | */ | |
89278c9d | 388 | fastopen: |
9092a76d JB |
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 | } | |
c10d9310 | 403 | NET_INC_STATS(sock_net(sk), |
9092a76d | 404 | LINUX_MIB_TCPFASTOPENPASSIVEFAIL); |
843f4a55 | 405 | } |
9092a76d | 406 | } |
7f9b838b | 407 | valid_foc.exp = foc->exp; |
89278c9d | 408 | *foc = valid_foc; |
7c85af88 | 409 | return NULL; |
5b7ed089 | 410 | } |
065263f4 WW |
411 | |
412 | bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss, | |
413 | struct tcp_fastopen_cookie *cookie) | |
414 | { | |
71c02379 | 415 | const struct dst_entry *dst; |
065263f4 | 416 | |
7268586b | 417 | tcp_fastopen_cache_get(sk, mss, cookie); |
cf1ef3f0 WW |
418 | |
419 | /* Firewall blackhole issue check */ | |
420 | if (tcp_fastopen_active_should_disable(sk)) { | |
421 | cookie->len = -1; | |
422 | return false; | |
423 | } | |
424 | ||
71c02379 CP |
425 | dst = __sk_dst_get(sk); |
426 | ||
427 | if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) { | |
065263f4 WW |
428 | cookie->len = -1; |
429 | return true; | |
430 | } | |
48027478 JB |
431 | if (cookie->len > 0) |
432 | return true; | |
433 | tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE; | |
434 | return false; | |
065263f4 | 435 | } |
19f6d3f3 WW |
436 | |
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 | */ | |
444 | bool 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 | } | |
468 | EXPORT_SYMBOL(tcp_fastopen_defer_connect); | |
cf1ef3f0 WW |
469 | |
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 | |
7268586b YC |
478 | * 3. client side TFO socket has timed out three times consecutively during |
479 | * or after handshake | |
cf1ef3f0 WW |
480 | * 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 | ||
cf1ef3f0 WW |
486 | /* Disable active TFO and record current jiffies and |
487 | * tfo_active_disable_times | |
488 | */ | |
46c2fa39 | 489 | void tcp_fastopen_active_disable(struct sock *sk) |
cf1ef3f0 | 490 | { |
3733be14 | 491 | struct net *net = sock_net(sk); |
cf1ef3f0 | 492 | |
021266ec | 493 | if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout)) |
213ad73d WW |
494 | return; |
495 | ||
6f20c8ad ED |
496 | /* 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(); | |
3733be14 | 503 | atomic_inc(&net->ipv4.tfo_active_disable_times); |
6f20c8ad | 504 | |
3733be14 | 505 | NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE); |
cf1ef3f0 WW |
506 | } |
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 | */ | |
512 | bool tcp_fastopen_active_should_disable(struct sock *sk) | |
513 | { | |
021266ec KI |
514 | unsigned int tfo_bh_timeout = |
515 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout); | |
cf1ef3f0 | 516 | unsigned long timeout; |
213ad73d | 517 | int tfo_da_times; |
3733be14 | 518 | int multiplier; |
cf1ef3f0 | 519 | |
213ad73d WW |
520 | if (!tfo_bh_timeout) |
521 | return false; | |
522 | ||
523 | tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times); | |
cf1ef3f0 WW |
524 | if (!tfo_da_times) |
525 | return false; | |
526 | ||
6f20c8ad ED |
527 | /* Paired with smp_mb__before_atomic() in tcp_fastopen_active_disable() */ |
528 | smp_rmb(); | |
529 | ||
974d8f86 | 530 | /* Limit timeout to max: 2^6 * initial timeout */ |
cf1ef3f0 | 531 | multiplier = 1 << min(tfo_da_times - 1, 6); |
6f20c8ad ED |
532 | |
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)) | |
cf1ef3f0 WW |
537 | 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 | */ | |
552 | void tcp_fastopen_active_disable_ofo_check(struct sock *sk) | |
553 | { | |
554 | struct tcp_sock *tp = tcp_sk(sk); | |
cf1ef3f0 | 555 | struct dst_entry *dst; |
18a4c0ea | 556 | struct sk_buff *skb; |
cf1ef3f0 WW |
557 | |
558 | if (!tp->syn_fastopen) | |
559 | return; | |
560 | ||
561 | if (!tp->data_segs_in) { | |
18a4c0ea ED |
562 | skb = skb_rb_first(&tp->out_of_order_queue); |
563 | if (skb && !skb_rb_next(skb)) { | |
cf1ef3f0 | 564 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { |
46c2fa39 | 565 | tcp_fastopen_active_disable(sk); |
cf1ef3f0 WW |
566 | return; |
567 | } | |
568 | } | |
569 | } else if (tp->syn_fastopen_ch && | |
3733be14 | 570 | atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) { |
cf1ef3f0 WW |
571 | dst = sk_dst_get(sk); |
572 | if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK))) | |
3733be14 | 573 | atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0); |
cf1ef3f0 WW |
574 | dst_release(dst); |
575 | } | |
576 | } | |
7268586b YC |
577 | |
578 | void 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 | } |