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  *  $Id: syncookies.c,v 1.18 2002/02/01 22:01:04 davem Exp $
  13  */
  14
  15 #include <linux/tcp.h>
  16 #include <linux/slab.h>
  17 #include <linux/random.h>
  18 #include <linux/cryptohash.h>
  19 #include <linux/kernel.h>
  20 #include <net/tcp.h>
  21
  22 /* Timestamps: lowest 9 bits store TCP options */
  23 #define TSBITS 9
  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, cookie_scratch)[16 + 5 + SHA_WORKSPACE_WORDS];
  42
  43 static u32 cookie_hash(__be32 saddr, __be32 daddr, __be16 sport, __be16 dport,
  44                        u32 count, int c)
  45 {
  46         __u32 *tmp = __get_cpu_var(cookie_scratch);
  47
  48         memcpy(tmp + 4, syncookie_secret[c], sizeof(syncookie_secret[c]));
  49         tmp[0] = (__force u32)saddr;
  50         tmp[1] = (__force u32)daddr;
  51         tmp[2] = ((__force u32)sport << 16) + (__force u32)dport;
  52         tmp[3] = count;
  53         sha_transform(tmp + 16, (__u8 *)tmp, tmp + 16 + 5);
  54
  55         return tmp[17];
  56 }
  57
  58
  59 /*
  60  * when syncookies are in effect and tcp timestamps are enabled we encode
  61  * tcp options in the lowest 9 bits of the timestamp value that will be
  62  * sent in the syn-ack.
  63  * Since subsequent timestamps use the normal tcp_time_stamp value, we
  64  * must make sure that the resulting initial timestamp is <= tcp_time_stamp.
  65  */
  66 __u32 cookie_init_timestamp(struct request_sock *req)
  67 {
  68         struct inet_request_sock *ireq;
  69         u32 ts, ts_now = tcp_time_stamp;
  70         u32 options = 0;
  71
  72         ireq = inet_rsk(req);
  73         if (ireq->wscale_ok) {
  74                 options = ireq->snd_wscale;
  75                 options |= ireq->rcv_wscale << 4;
  76         }
  77         options |= ireq->sack_ok << 8;
  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  * This table has to be sorted and terminated with (__u16)-1.
 142  * XXX generate a better table.
 143  * Unresolved Issues: HIPPI with a 64k MSS is not well supported.
 144  */
 145 static __u16 const msstab[] = {
 146         64 - 1,
 147         256 - 1,
 148         512 - 1,
 149         536 - 1,
 150         1024 - 1,
 151         1440 - 1,
 152         1460 - 1,
 153         4312 - 1,
 154         (__u16)-1
 155 };
 156 /* The number doesn't include the -1 terminator */
 157 #define NUM_MSS (ARRAY_SIZE(msstab) - 1)
 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         struct tcp_sock *tp = tcp_sk(sk);
 166         const struct iphdr *iph = ip_hdr(skb);
 167         const struct tcphdr *th = tcp_hdr(skb);
 168         int mssind;
 169         const __u16 mss = *mssp;
 170
 171         tp->last_synq_overflow = jiffies;
 172
 173         /* XXX sort msstab[] by probability?  Binary search? */
 174         for (mssind = 0; mss > msstab[mssind + 1]; mssind++)
 175                 ;
 176         *mssp = msstab[mssind] + 1;
 177
 178         NET_INC_STATS_BH(LINUX_MIB_SYNCOOKIESSENT);
 179
 180         return secure_tcp_syn_cookie(iph->saddr, iph->daddr,
 181                                      th->source, th->dest, ntohl(th->seq),
 182                                      jiffies / (HZ * 60), mssind);
 183 }
 184
 185 /*
 186  * This (misnamed) value is the age of syncookie which is permitted.
 187  * Its ideal value should be dependent on TCP_TIMEOUT_INIT and
 188  * sysctl_tcp_retries1. It's a rather complicated formula (exponential
 189  * backoff) to compute at runtime so it's currently hardcoded here.
 190  */
 191 #define COUNTER_TRIES 4
 192 /*
 193  * Check if a ack sequence number is a valid syncookie.
 194  * Return the decoded mss if it is, or 0 if not.
 195  */
 196 static inline int cookie_check(struct sk_buff *skb, __u32 cookie)
 197 {
 198         const struct iphdr *iph = ip_hdr(skb);
 199         const struct tcphdr *th = tcp_hdr(skb);
 200         __u32 seq = ntohl(th->seq) - 1;
 201         __u32 mssind = check_tcp_syn_cookie(cookie, iph->saddr, iph->daddr,
 202                                             th->source, th->dest, seq,
 203                                             jiffies / (HZ * 60),
 204                                             COUNTER_TRIES);
 205
 206         return mssind < NUM_MSS ? msstab[mssind] + 1 : 0;
 207 }
 208
 209 static inline struct sock *get_cookie_sock(struct sock *sk, struct sk_buff *skb,
 210                                            struct request_sock *req,
 211                                            struct dst_entry *dst)
 212 {
 213         struct inet_connection_sock *icsk = inet_csk(sk);
 214         struct sock *child;
 215
 216         child = icsk->icsk_af_ops->syn_recv_sock(sk, skb, req, dst);
 217         if (child)
 218                 inet_csk_reqsk_queue_add(sk, req, child);
 219         else
 220                 reqsk_free(req);
 221
 222         return child;
 223 }
 224
 225
 226 /*
 227  * when syncookies are in effect and tcp timestamps are enabled we stored
 228  * additional tcp options in the timestamp.
 229  * This extracts these options from the timestamp echo.
 230  *
 231  * The lowest 4 bits are for snd_wscale
 232  * The next 4 lsb are for rcv_wscale
 233  * The next lsb is for sack_ok
 234  */
 235 void cookie_check_timestamp(struct tcp_options_received *tcp_opt)
 236 {
 237         /* echoed timestamp, 9 lowest bits contain options */
 238         u32 options = tcp_opt->rcv_tsecr & TSMASK;
 239
 240         tcp_opt->snd_wscale = options & 0xf;
 241         options >>= 4;
 242         tcp_opt->rcv_wscale = options & 0xf;
 243
 244         tcp_opt->sack_ok = (options >> 4) & 0x1;
 245
 246         if (tcp_opt->sack_ok)
 247                 tcp_sack_reset(tcp_opt);
 248
 249         if (tcp_opt->snd_wscale || tcp_opt->rcv_wscale)
 250                 tcp_opt->wscale_ok = 1;
 251 }
 252 EXPORT_SYMBOL(cookie_check_timestamp);
 253
 254 struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb,
 255                              struct ip_options *opt)
 256 {
 257         struct inet_request_sock *ireq;
 258         struct tcp_request_sock *treq;
 259         struct tcp_sock *tp = tcp_sk(sk);
 260         const struct tcphdr *th = tcp_hdr(skb);
 261         __u32 cookie = ntohl(th->ack_seq) - 1;
 262         struct sock *ret = sk;
 263         struct request_sock *req;
 264         int mss;
 265         struct rtable *rt;
 266         __u8 rcv_wscale;
 267         struct tcp_options_received tcp_opt;
 268
 269         if (!sysctl_tcp_syncookies || !th->ack)
 270                 goto out;
 271
 272         if (time_after(jiffies, tp->last_synq_overflow + TCP_TIMEOUT_INIT) ||
 273             (mss = cookie_check(skb, cookie)) == 0) {
 274                 NET_INC_STATS_BH(LINUX_MIB_SYNCOOKIESFAILED);
 275                 goto out;
 276         }
 277
 278         NET_INC_STATS_BH(LINUX_MIB_SYNCOOKIESRECV);
 279
 280         /* check for timestamp cookie support */
 281         memset(&tcp_opt, 0, sizeof(tcp_opt));
 282         tcp_parse_options(skb, &tcp_opt, 0);
 283
 284         if (tcp_opt.saw_tstamp)
 285                 cookie_check_timestamp(&tcp_opt);
 286
 287         ret = NULL;
 288         req = reqsk_alloc(&tcp_request_sock_ops); /* for safety */
 289         if (!req)
 290                 goto out;
 291
 292         if (security_inet_conn_request(sk, skb, req)) {
 293                 reqsk_free(req);
 294                 goto out;
 295         }
 296         ireq = inet_rsk(req);
 297         treq = tcp_rsk(req);
 298         treq->rcv_isn           = ntohl(th->seq) - 1;
 299         treq->snt_isn           = cookie;
 300         req->mss                = mss;
 301         ireq->rmt_port          = th->source;
 302         ireq->loc_addr          = ip_hdr(skb)->daddr;
 303         ireq->rmt_addr          = ip_hdr(skb)->saddr;
 304         ireq->opt               = NULL;
 305         ireq->snd_wscale        = tcp_opt.snd_wscale;
 306         ireq->rcv_wscale        = tcp_opt.rcv_wscale;
 307         ireq->sack_ok           = tcp_opt.sack_ok;
 308         ireq->wscale_ok         = tcp_opt.wscale_ok;
 309         ireq->tstamp_ok         = tcp_opt.saw_tstamp;
 310         req->ts_recent          = tcp_opt.saw_tstamp ? tcp_opt.rcv_tsval : 0;
 311
 312         /* We throwed the options of the initial SYN away, so we hope
 313          * the ACK carries the same options again (see RFC1122 4.2.3.8)
 314          */
 315         if (opt && opt->optlen) {
 316                 int opt_size = sizeof(struct ip_options) + opt->optlen;
 317
 318                 ireq->opt = kmalloc(opt_size, GFP_ATOMIC);
 319                 if (ireq->opt != NULL && ip_options_echo(ireq->opt, skb)) {
 320                         kfree(ireq->opt);
 321                         ireq->opt = NULL;
 322                 }
 323         }
 324
 325         req->expires    = 0UL;
 326         req->retrans    = 0;
 327
 328         /*
 329          * We need to lookup the route here to get at the correct
 330          * window size. We should better make sure that the window size
 331          * hasn't changed since we received the original syn, but I see
 332          * no easy way to do this.
 333          */
 334         {
 335                 struct flowi fl = { .nl_u = { .ip4_u =
 336                                               { .daddr = ((opt && opt->srr) ?
 337                                                           opt->faddr :
 338                                                           ireq->rmt_addr),
 339                                                 .saddr = ireq->loc_addr,
 340                                                 .tos = RT_CONN_FLAGS(sk) } },
 341                                     .proto = IPPROTO_TCP,
 342                                     .uli_u = { .ports =
 343                                                { .sport = th->dest,
 344                                                  .dport = th->source } } };
 345                 security_req_classify_flow(req, &fl);
 346                 if (ip_route_output_key(&init_net, &rt, &fl)) {
 347                         reqsk_free(req);
 348                         goto out;
 349                 }
 350         }
 351
 352         /* Try to redo what tcp_v4_send_synack did. */
 353         req->window_clamp = tp->window_clamp ? :dst_metric(&rt->u.dst, RTAX_WINDOW);
 354
 355         tcp_select_initial_window(tcp_full_space(sk), req->mss,
 356                                   &req->rcv_wnd, &req->window_clamp,
 357                                   ireq->wscale_ok, &rcv_wscale);
 358
 359         ireq->rcv_wscale  = rcv_wscale;
 360
 361         ret = get_cookie_sock(sk, skb, req, &rt->u.dst);
 362 out:    return ret;
 363 }