]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blob - net/ipv4/syncookies.c
tcp: TCP Fast Open Server - support TFO listeners
[mirror_ubuntu-zesty-kernel.git] / net / ipv4 / syncookies.c
1 /*
2 * Syncookies implementation for the Linux kernel
3 *
4 * Copyright (C) 1997 Andi Kleen
5 * Based on ideas by D.J.Bernstein and Eric Schenk.
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version
10 * 2 of the License, or (at your option) any later version.
11 */
12
13 #include <linux/tcp.h>
14 #include <linux/slab.h>
15 #include <linux/random.h>
16 #include <linux/cryptohash.h>
17 #include <linux/kernel.h>
18 #include <linux/export.h>
19 #include <net/tcp.h>
20 #include <net/route.h>
21
22 /* Timestamps: lowest bits store TCP options */
23 #define TSBITS 6
24 #define TSMASK (((__u32)1 << TSBITS) - 1)
25
26 extern int sysctl_tcp_syncookies;
27
28 __u32 syncookie_secret[2][16-4+SHA_DIGEST_WORDS];
29 EXPORT_SYMBOL(syncookie_secret);
30
31 static __init int init_syncookies(void)
32 {
33 get_random_bytes(syncookie_secret, sizeof(syncookie_secret));
34 return 0;
35 }
36 __initcall(init_syncookies);
37
38 #define COOKIEBITS 24 /* Upper bits store count */
39 #define COOKIEMASK (((__u32)1 << COOKIEBITS) - 1)
40
41 static DEFINE_PER_CPU(__u32 [16 + 5 + SHA_WORKSPACE_WORDS],
42 ipv4_cookie_scratch);
43
44 static u32 cookie_hash(__be32 saddr, __be32 daddr, __be16 sport, __be16 dport,
45 u32 count, int c)
46 {
47 __u32 *tmp = __get_cpu_var(ipv4_cookie_scratch);
48
49 memcpy(tmp + 4, syncookie_secret[c], sizeof(syncookie_secret[c]));
50 tmp[0] = (__force u32)saddr;
51 tmp[1] = (__force u32)daddr;
52 tmp[2] = ((__force u32)sport << 16) + (__force u32)dport;
53 tmp[3] = count;
54 sha_transform(tmp + 16, (__u8 *)tmp, tmp + 16 + 5);
55
56 return tmp[17];
57 }
58
59
60 /*
61 * when syncookies are in effect and tcp timestamps are enabled we encode
62 * tcp options in the lower bits of the timestamp value that will be
63 * sent in the syn-ack.
64 * Since subsequent timestamps use the normal tcp_time_stamp value, we
65 * must make sure that the resulting initial timestamp is <= tcp_time_stamp.
66 */
67 __u32 cookie_init_timestamp(struct request_sock *req)
68 {
69 struct inet_request_sock *ireq;
70 u32 ts, ts_now = tcp_time_stamp;
71 u32 options = 0;
72
73 ireq = inet_rsk(req);
74
75 options = ireq->wscale_ok ? ireq->snd_wscale : 0xf;
76 options |= ireq->sack_ok << 4;
77 options |= ireq->ecn_ok << 5;
78
79 ts = ts_now & ~TSMASK;
80 ts |= options;
81 if (ts > ts_now) {
82 ts >>= TSBITS;
83 ts--;
84 ts <<= TSBITS;
85 ts |= options;
86 }
87 return ts;
88 }
89
90
91 static __u32 secure_tcp_syn_cookie(__be32 saddr, __be32 daddr, __be16 sport,
92 __be16 dport, __u32 sseq, __u32 count,
93 __u32 data)
94 {
95 /*
96 * Compute the secure sequence number.
97 * The output should be:
98 * HASH(sec1,saddr,sport,daddr,dport,sec1) + sseq + (count * 2^24)
99 * + (HASH(sec2,saddr,sport,daddr,dport,count,sec2) % 2^24).
100 * Where sseq is their sequence number and count increases every
101 * minute by 1.
102 * As an extra hack, we add a small "data" value that encodes the
103 * MSS into the second hash value.
104 */
105
106 return (cookie_hash(saddr, daddr, sport, dport, 0, 0) +
107 sseq + (count << COOKIEBITS) +
108 ((cookie_hash(saddr, daddr, sport, dport, count, 1) + data)
109 & COOKIEMASK));
110 }
111
112 /*
113 * This retrieves the small "data" value from the syncookie.
114 * If the syncookie is bad, the data returned will be out of
115 * range. This must be checked by the caller.
116 *
117 * The count value used to generate the cookie must be within
118 * "maxdiff" if the current (passed-in) "count". The return value
119 * is (__u32)-1 if this test fails.
120 */
121 static __u32 check_tcp_syn_cookie(__u32 cookie, __be32 saddr, __be32 daddr,
122 __be16 sport, __be16 dport, __u32 sseq,
123 __u32 count, __u32 maxdiff)
124 {
125 __u32 diff;
126
127 /* Strip away the layers from the cookie */
128 cookie -= cookie_hash(saddr, daddr, sport, dport, 0, 0) + sseq;
129
130 /* Cookie is now reduced to (count * 2^24) ^ (hash % 2^24) */
131 diff = (count - (cookie >> COOKIEBITS)) & ((__u32) - 1 >> COOKIEBITS);
132 if (diff >= maxdiff)
133 return (__u32)-1;
134
135 return (cookie -
136 cookie_hash(saddr, daddr, sport, dport, count - diff, 1))
137 & COOKIEMASK; /* Leaving the data behind */
138 }
139
140 /*
141 * MSS Values are taken from the 2009 paper
142 * 'Measuring TCP Maximum Segment Size' by S. Alcock and R. Nelson:
143 * - values 1440 to 1460 accounted for 80% of observed mss values
144 * - values outside the 536-1460 range are rare (<0.2%).
145 *
146 * Table must be sorted.
147 */
148 static __u16 const msstab[] = {
149 64,
150 512,
151 536,
152 1024,
153 1440,
154 1460,
155 4312,
156 8960,
157 };
158
159 /*
160 * Generate a syncookie. mssp points to the mss, which is returned
161 * rounded down to the value encoded in the cookie.
162 */
163 __u32 cookie_v4_init_sequence(struct sock *sk, struct sk_buff *skb, __u16 *mssp)
164 {
165 const struct iphdr *iph = ip_hdr(skb);
166 const struct tcphdr *th = tcp_hdr(skb);
167 int mssind;
168 const __u16 mss = *mssp;
169
170 tcp_synq_overflow(sk);
171
172 for (mssind = ARRAY_SIZE(msstab) - 1; mssind ; mssind--)
173 if (mss >= msstab[mssind])
174 break;
175 *mssp = msstab[mssind];
176
177 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SYNCOOKIESSENT);
178
179 return secure_tcp_syn_cookie(iph->saddr, iph->daddr,
180 th->source, th->dest, ntohl(th->seq),
181 jiffies / (HZ * 60), mssind);
182 }
183
184 /*
185 * This (misnamed) value is the age of syncookie which is permitted.
186 * Its ideal value should be dependent on TCP_TIMEOUT_INIT and
187 * sysctl_tcp_retries1. It's a rather complicated formula (exponential
188 * backoff) to compute at runtime so it's currently hardcoded here.
189 */
190 #define COUNTER_TRIES 4
191 /*
192 * Check if a ack sequence number is a valid syncookie.
193 * Return the decoded mss if it is, or 0 if not.
194 */
195 static inline int cookie_check(struct sk_buff *skb, __u32 cookie)
196 {
197 const struct iphdr *iph = ip_hdr(skb);
198 const struct tcphdr *th = tcp_hdr(skb);
199 __u32 seq = ntohl(th->seq) - 1;
200 __u32 mssind = check_tcp_syn_cookie(cookie, iph->saddr, iph->daddr,
201 th->source, th->dest, seq,
202 jiffies / (HZ * 60),
203 COUNTER_TRIES);
204
205 return mssind < ARRAY_SIZE(msstab) ? msstab[mssind] : 0;
206 }
207
208 static inline struct sock *get_cookie_sock(struct sock *sk, struct sk_buff *skb,
209 struct request_sock *req,
210 struct dst_entry *dst)
211 {
212 struct inet_connection_sock *icsk = inet_csk(sk);
213 struct sock *child;
214
215 child = icsk->icsk_af_ops->syn_recv_sock(sk, skb, req, dst);
216 if (child)
217 inet_csk_reqsk_queue_add(sk, req, child);
218 else
219 reqsk_free(req);
220
221 return child;
222 }
223
224
225 /*
226 * when syncookies are in effect and tcp timestamps are enabled we stored
227 * additional tcp options in the timestamp.
228 * This extracts these options from the timestamp echo.
229 *
230 * The lowest 4 bits store snd_wscale.
231 * next 2 bits indicate SACK and ECN support.
232 *
233 * return false if we decode an option that should not be.
234 */
235 bool cookie_check_timestamp(struct tcp_options_received *tcp_opt, bool *ecn_ok)
236 {
237 /* echoed timestamp, lowest bits contain options */
238 u32 options = tcp_opt->rcv_tsecr & TSMASK;
239
240 if (!tcp_opt->saw_tstamp) {
241 tcp_clear_options(tcp_opt);
242 return true;
243 }
244
245 if (!sysctl_tcp_timestamps)
246 return false;
247
248 tcp_opt->sack_ok = (options & (1 << 4)) ? TCP_SACK_SEEN : 0;
249 *ecn_ok = (options >> 5) & 1;
250 if (*ecn_ok && !sysctl_tcp_ecn)
251 return false;
252
253 if (tcp_opt->sack_ok && !sysctl_tcp_sack)
254 return false;
255
256 if ((options & 0xf) == 0xf)
257 return true; /* no window scaling */
258
259 tcp_opt->wscale_ok = 1;
260 tcp_opt->snd_wscale = options & 0xf;
261 return sysctl_tcp_window_scaling != 0;
262 }
263 EXPORT_SYMBOL(cookie_check_timestamp);
264
265 struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb,
266 struct ip_options *opt)
267 {
268 struct tcp_options_received tcp_opt;
269 const u8 *hash_location;
270 struct inet_request_sock *ireq;
271 struct tcp_request_sock *treq;
272 struct tcp_sock *tp = tcp_sk(sk);
273 const struct tcphdr *th = tcp_hdr(skb);
274 __u32 cookie = ntohl(th->ack_seq) - 1;
275 struct sock *ret = sk;
276 struct request_sock *req;
277 int mss;
278 struct rtable *rt;
279 __u8 rcv_wscale;
280 bool ecn_ok = false;
281 struct flowi4 fl4;
282
283 if (!sysctl_tcp_syncookies || !th->ack || th->rst)
284 goto out;
285
286 if (tcp_synq_no_recent_overflow(sk) ||
287 (mss = cookie_check(skb, cookie)) == 0) {
288 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SYNCOOKIESFAILED);
289 goto out;
290 }
291
292 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SYNCOOKIESRECV);
293
294 /* check for timestamp cookie support */
295 memset(&tcp_opt, 0, sizeof(tcp_opt));
296 tcp_parse_options(skb, &tcp_opt, &hash_location, 0, NULL);
297
298 if (!cookie_check_timestamp(&tcp_opt, &ecn_ok))
299 goto out;
300
301 ret = NULL;
302 req = inet_reqsk_alloc(&tcp_request_sock_ops); /* for safety */
303 if (!req)
304 goto out;
305
306 ireq = inet_rsk(req);
307 treq = tcp_rsk(req);
308 treq->rcv_isn = ntohl(th->seq) - 1;
309 treq->snt_isn = cookie;
310 req->mss = mss;
311 ireq->loc_port = th->dest;
312 ireq->rmt_port = th->source;
313 ireq->loc_addr = ip_hdr(skb)->daddr;
314 ireq->rmt_addr = ip_hdr(skb)->saddr;
315 ireq->ecn_ok = ecn_ok;
316 ireq->snd_wscale = tcp_opt.snd_wscale;
317 ireq->sack_ok = tcp_opt.sack_ok;
318 ireq->wscale_ok = tcp_opt.wscale_ok;
319 ireq->tstamp_ok = tcp_opt.saw_tstamp;
320 req->ts_recent = tcp_opt.saw_tstamp ? tcp_opt.rcv_tsval : 0;
321 treq->snt_synack = tcp_opt.saw_tstamp ? tcp_opt.rcv_tsecr : 0;
322 treq->listener = NULL;
323
324 /* We throwed the options of the initial SYN away, so we hope
325 * the ACK carries the same options again (see RFC1122 4.2.3.8)
326 */
327 if (opt && opt->optlen) {
328 int opt_size = sizeof(struct ip_options_rcu) + opt->optlen;
329
330 ireq->opt = kmalloc(opt_size, GFP_ATOMIC);
331 if (ireq->opt != NULL && ip_options_echo(&ireq->opt->opt, skb)) {
332 kfree(ireq->opt);
333 ireq->opt = NULL;
334 }
335 }
336
337 if (security_inet_conn_request(sk, skb, req)) {
338 reqsk_free(req);
339 goto out;
340 }
341
342 req->expires = 0UL;
343 req->retrans = 0;
344
345 /*
346 * We need to lookup the route here to get at the correct
347 * window size. We should better make sure that the window size
348 * hasn't changed since we received the original syn, but I see
349 * no easy way to do this.
350 */
351 flowi4_init_output(&fl4, 0, sk->sk_mark, RT_CONN_FLAGS(sk),
352 RT_SCOPE_UNIVERSE, IPPROTO_TCP,
353 inet_sk_flowi_flags(sk),
354 (opt && opt->srr) ? opt->faddr : ireq->rmt_addr,
355 ireq->loc_addr, th->source, th->dest);
356 security_req_classify_flow(req, flowi4_to_flowi(&fl4));
357 rt = ip_route_output_key(sock_net(sk), &fl4);
358 if (IS_ERR(rt)) {
359 reqsk_free(req);
360 goto out;
361 }
362
363 /* Try to redo what tcp_v4_send_synack did. */
364 req->window_clamp = tp->window_clamp ? :dst_metric(&rt->dst, RTAX_WINDOW);
365
366 tcp_select_initial_window(tcp_full_space(sk), req->mss,
367 &req->rcv_wnd, &req->window_clamp,
368 ireq->wscale_ok, &rcv_wscale,
369 dst_metric(&rt->dst, RTAX_INITRWND));
370
371 ireq->rcv_wscale = rcv_wscale;
372
373 ret = get_cookie_sock(sk, skb, req, &rt->dst);
374 /* ip_queue_xmit() depends on our flow being setup
375 * Normal sockets get it right from inet_csk_route_child_sock()
376 */
377 if (ret)
378 inet_sk(ret)->cork.fl.u.ip4 = fl4;
379 out: return ret;
380 }