sys/netinet/tcp_fastopen.c

   1 /*-
   2  * Copyright (c) 2015-2017 Patrick Kelsey
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 /*
  28  * This is an implementation of TCP Fast Open (TFO) [RFC7413]. To include
  29  * this code, add the following line to your kernel config:
  30  *
  31  * options TCP_RFC7413
  32  *
  33  *
  34  * The generated TFO cookies are the 64-bit output of
  35  * SipHash24(key=<16-byte-key>, msg=<client-ip>).  Multiple concurrent valid
  36  * keys are supported so that time-based rolling cookie invalidation
  37  * policies can be implemented in the system.  The default number of
  38  * concurrent keys is 2.  This can be adjusted in the kernel config as
  39  * follows:
  40  *
  41  * options TCP_RFC7413_MAX_KEYS=<num-keys>
  42  *
  43  *
  44  * In addition to the facilities defined in RFC7413, this implementation
  45  * supports a pre-shared key (PSK) mode of operation in which the TFO server
  46  * requires the client to be in posession of a shared secret in order for
  47  * the client to be able to successfully open TFO connections with the
  48  * server.  This is useful, for example, in environments where TFO servers
  49  * are exposed to both internal and external clients and only wish to allow
  50  * TFO connections from internal clients.
  51  *
  52  * In the PSK mode of operation, the server generates and sends TFO cookies
  53  * to requesting clients as usual.  However, when validating cookies
  54  * received in TFO SYNs from clients, the server requires the
  55  * client-supplied cookie to equal SipHash24(key=<16-byte-psk>,
  56  * msg=<cookie-sent-to-client>).
  57  *
  58  * Multiple concurrent valid pre-shared keys are supported so that
  59  * time-based rolling PSK invalidation policies can be implemented in the
  60  * system.  The default number of concurrent pre-shared keys is 2.  This can
  61  * be adjusted in the kernel config as follows:
  62  *
  63  * options TCP_RFC7413_MAX_PSKS=<num-psks>
  64  *
  65  *
  66  * The following TFO-specific sysctls are defined:
  67  *
  68  * net.inet.tcp.fastopen.acceptany (RW, default 0)
  69  *     When non-zero, all client-supplied TFO cookies will be considered to
  70  *     be valid.
  71  *
  72  * net.inet.tcp.fastopen.autokey (RW, default 120)
  73  *     When this and net.inet.tcp.fastopen.server_enable are non-zero, a new
  74  *     key will be automatically generated after this many seconds.
  75  *
  76  * net.inet.tcp.fastopen.ccache_bucket_limit
  77  *                     (RWTUN, default TCP_FASTOPEN_CCACHE_BUCKET_LIMIT_DEFAULT)
  78  *     The maximum number of entries in a client cookie cache bucket.
  79  *
  80  * net.inet.tcp.fastopen.ccache_buckets
  81  *                          (RDTUN, default TCP_FASTOPEN_CCACHE_BUCKETS_DEFAULT)
  82  *     The number of client cookie cache buckets.
  83  *
  84  * net.inet.tcp.fastopen.ccache_list (RO)
  85  *     Print the client cookie cache.
  86  *
  87  * net.inet.tcp.fastopen.client_enable (RW, default 0)
  88  *     When zero, no new active (i.e., client) TFO connections can be
  89  *     created.  On the transition from enabled to disabled, the client
  90  *     cookie cache is cleared and disabled.  The transition from enabled to
  91  *     disabled does not affect any active TFO connections in progress; it
  92  *     only prevents new ones from being made.
  93  *
  94  * net.inet.tcp.fastopen.keylen (RD)
  95  *     The key length in bytes.
  96  *
  97  * net.inet.tcp.fastopen.maxkeys (RD)
  98  *     The maximum number of keys supported.
  99  *
 100  * net.inet.tcp.fastopen.maxpsks (RD)
 101  *     The maximum number of pre-shared keys supported.
 102  *
 103  * net.inet.tcp.fastopen.numkeys (RD)
 104  *     The current number of keys installed.
 105  *
 106  * net.inet.tcp.fastopen.numpsks (RD)
 107  *     The current number of pre-shared keys installed.
 108  *
 109  * net.inet.tcp.fastopen.path_disable_time
 110  *                          (RW, default TCP_FASTOPEN_PATH_DISABLE_TIME_DEFAULT)
 111  *     When a failure occurs while trying to create a new active (i.e.,
 112  *     client) TFO connection, new active connections on the same path, as
 113  *     determined by the tuple {client_ip, server_ip, server_port}, will be
 114  *     forced to be non-TFO for this many seconds.  Note that the path
 115  *     disable mechanism relies on state stored in client cookie cache
 116  *     entries, so it is possible for the disable time for a given path to
 117  *     be reduced if the corresponding client cookie cache entry is reused
 118  *     due to resource pressure before the disable period has elapsed.
 119  *
 120  * net.inet.tcp.fastopen.psk_enable (RW, default 0)
 121  *     When non-zero, pre-shared key (PSK) mode is enabled for all TFO
 122  *     servers.  On the transition from enabled to disabled, all installed
 123  *     pre-shared keys are removed.
 124  *
 125  * net.inet.tcp.fastopen.server_enable (RW, default 0)
 126  *     When zero, no new passive (i.e., server) TFO connections can be
 127  *     created.  On the transition from enabled to disabled, all installed
 128  *     keys and pre-shared keys are removed.  On the transition from
 129  *     disabled to enabled, if net.inet.tcp.fastopen.autokey is non-zero and
 130  *     there are no keys installed, a new key will be generated immediately.
 131  *     The transition from enabled to disabled does not affect any passive
 132  *     TFO connections in progress; it only prevents new ones from being
 133  *     made.
 134  *
 135  * net.inet.tcp.fastopen.setkey (WR)
 136  *     Install a new key by writing net.inet.tcp.fastopen.keylen bytes to
 137  *     this sysctl.
 138  *
 139  * net.inet.tcp.fastopen.setpsk (WR)
 140  *     Install a new pre-shared key by writing net.inet.tcp.fastopen.keylen
 141  *     bytes to this sysctl.
 142  *
 143  * In order for TFO connections to be created via a listen socket, that
 144  * socket must have the TCP_FASTOPEN socket option set on it.  This option
 145  * can be set on the socket either before or after the listen() is invoked.
 146  * Clearing this option on a listen socket after it has been set has no
 147  * effect on existing TFO connections or TFO connections in progress; it
 148  * only prevents new TFO connections from being made.
 149  *
 150  * For passively-created sockets, the TCP_FASTOPEN socket option can be
 151  * queried to determine whether the connection was established using TFO.
 152  * Note that connections that are established via a TFO SYN, but that fall
 153  * back to using a non-TFO SYN|ACK will have the TCP_FASTOPEN socket option
 154  * set.
 155  *
 156  * Per the RFC, this implementation limits the number of TFO connections
 157  * that can be in the SYN_RECEIVED state on a per listen-socket basis.
 158  * Whenever this limit is exceeded, requests for new TFO connections are
 159  * serviced as non-TFO requests.  Without such a limit, given a valid TFO
 160  * cookie, an attacker could keep the listen queue in an overflow condition
 161  * using a TFO SYN flood.  This implementation sets the limit at half the
 162  * configured listen backlog.
 163  *
 164  */
 165
 166 #include <sys/cdefs.h>
 167 __FBSDID("$FreeBSD$");
 168
 169 #include "opt_inet.h"
 170
 171 #include <sys/param.h>
 172 #include <sys/jail.h>
 173 #include <sys/kernel.h>
 174 #include <sys/hash.h>
 175 #include <sys/limits.h>
 176 #include <sys/lock.h>
 177 #include <sys/proc.h>
 178 #include <sys/rmlock.h>
 179 #include <sys/sbuf.h>
 180 #include <sys/socket.h>
 181 #include <sys/socketvar.h>
 182 #include <sys/sysctl.h>
 183 #include <sys/systm.h>
 184
 185 #include <crypto/siphash/siphash.h>
 186
 187 #include <net/vnet.h>
 188
 189 #include <netinet/in.h>
 190 #include <netinet/in_pcb.h>
 191 #include <netinet/tcp_var.h>
 192 #include <netinet/tcp_fastopen.h>
 193
 194 #define TCP_FASTOPEN_KEY_LEN    SIPHASH_KEY_LENGTH
 195
 196 #if TCP_FASTOPEN_PSK_LEN != TCP_FASTOPEN_KEY_LEN
 197 #error TCP_FASTOPEN_PSK_LEN must be equal to TCP_FASTOPEN_KEY_LEN
 198 #endif
 199
 200 /*
 201  * Because a PSK-mode setsockopt() uses tcpcb.t_tfo_cookie.client to hold
 202  * the PSK until the connect occurs.
 203  */
 204 #if TCP_FASTOPEN_MAX_COOKIE_LEN < TCP_FASTOPEN_PSK_LEN
 205 #error TCP_FASTOPEN_MAX_COOKIE_LEN must be >= TCP_FASTOPEN_PSK_LEN
 206 #endif
 207
 208 #define TCP_FASTOPEN_CCACHE_BUCKET_LIMIT_DEFAULT        16
 209 #define TCP_FASTOPEN_CCACHE_BUCKETS_DEFAULT             2048 /* must be power of 2 */
 210
 211 #define TCP_FASTOPEN_PATH_DISABLE_TIME_DEFAULT          900 /* seconds */
 212
 213 #if !defined(TCP_RFC7413_MAX_KEYS) || (TCP_RFC7413_MAX_KEYS < 1)
 214 #define TCP_FASTOPEN_MAX_KEYS   2
 215 #else
 216 #define TCP_FASTOPEN_MAX_KEYS   TCP_RFC7413_MAX_KEYS
 217 #endif
 218
 219 #if TCP_FASTOPEN_MAX_KEYS > 10
 220 #undef TCP_FASTOPEN_MAX_KEYS
 221 #define TCP_FASTOPEN_MAX_KEYS   10
 222 #endif
 223
 224 #if !defined(TCP_RFC7413_MAX_PSKS) || (TCP_RFC7413_MAX_PSKS < 1)
 225 #define TCP_FASTOPEN_MAX_PSKS   2
 226 #else
 227 #define TCP_FASTOPEN_MAX_PSKS   TCP_RFC7413_MAX_PSKS
 228 #endif
 229
 230 #if TCP_FASTOPEN_MAX_PSKS > 10
 231 #undef TCP_FASTOPEN_MAX_PSKS
 232 #define TCP_FASTOPEN_MAX_PSKS   10
 233 #endif
 234
 235 struct tcp_fastopen_keylist {
 236         unsigned int newest;
 237         unsigned int newest_psk;
 238         uint8_t key[TCP_FASTOPEN_MAX_KEYS][TCP_FASTOPEN_KEY_LEN];
 239         uint8_t psk[TCP_FASTOPEN_MAX_PSKS][TCP_FASTOPEN_KEY_LEN];
 240 };
 241
 242 struct tcp_fastopen_callout {
 243         struct callout c;
 244         struct vnet *v;
 245 };
 246
 247 static struct tcp_fastopen_ccache_entry *tcp_fastopen_ccache_lookup(
 248     struct in_conninfo *, struct tcp_fastopen_ccache_bucket **);
 249 static struct tcp_fastopen_ccache_entry *tcp_fastopen_ccache_create(
 250     struct tcp_fastopen_ccache_bucket *, struct in_conninfo *, uint16_t, uint8_t,
 251     uint8_t *);
 252 static void tcp_fastopen_ccache_bucket_trim(struct tcp_fastopen_ccache_bucket *,
 253     unsigned int);
 254 static void tcp_fastopen_ccache_entry_drop(struct tcp_fastopen_ccache_entry *,
 255     struct tcp_fastopen_ccache_bucket *);
 256
 257 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, fastopen, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 258     "TCP Fast Open");
 259
 260 VNET_DEFINE_STATIC(int, tcp_fastopen_acceptany) = 0;
 261 #define V_tcp_fastopen_acceptany        VNET(tcp_fastopen_acceptany)
 262 SYSCTL_INT(_net_inet_tcp_fastopen, OID_AUTO, acceptany,
 263     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_fastopen_acceptany), 0,
 264     "Accept any non-empty cookie");
 265
 266 VNET_DEFINE_STATIC(unsigned int, tcp_fastopen_autokey) = 120;
 267 #define V_tcp_fastopen_autokey  VNET(tcp_fastopen_autokey)
 268 static int sysctl_net_inet_tcp_fastopen_autokey(SYSCTL_HANDLER_ARGS);
 269 SYSCTL_PROC(_net_inet_tcp_fastopen, OID_AUTO, autokey,
 270     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
 271     NULL, 0, &sysctl_net_inet_tcp_fastopen_autokey, "IU",
 272     "Number of seconds between auto-generation of a new key; zero disables");
 273
 274 static int sysctl_net_inet_tcp_fastopen_ccache_bucket_limit(SYSCTL_HANDLER_ARGS);
 275 SYSCTL_PROC(_net_inet_tcp_fastopen, OID_AUTO, ccache_bucket_limit,
 276     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT,
 277     NULL, 0, &sysctl_net_inet_tcp_fastopen_ccache_bucket_limit, "IU",
 278     "Max entries per bucket in client cookie cache");
 279
 280 VNET_DEFINE_STATIC(unsigned int, tcp_fastopen_ccache_buckets) =
 281     TCP_FASTOPEN_CCACHE_BUCKETS_DEFAULT;
 282 #define V_tcp_fastopen_ccache_buckets VNET(tcp_fastopen_ccache_buckets)
 283 SYSCTL_UINT(_net_inet_tcp_fastopen, OID_AUTO, ccache_buckets,
 284     CTLFLAG_VNET | CTLFLAG_RDTUN, &VNET_NAME(tcp_fastopen_ccache_buckets), 0,
 285     "Client cookie cache number of buckets (power of 2)");
 286
 287 VNET_DEFINE(unsigned int, tcp_fastopen_client_enable) = 1;
 288 static int sysctl_net_inet_tcp_fastopen_client_enable(SYSCTL_HANDLER_ARGS);
 289 SYSCTL_PROC(_net_inet_tcp_fastopen, OID_AUTO, client_enable,
 290     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 291     NULL, 0, &sysctl_net_inet_tcp_fastopen_client_enable, "IU",
 292     "Enable/disable TCP Fast Open client functionality");
 293
 294 SYSCTL_INT(_net_inet_tcp_fastopen, OID_AUTO, keylen,
 295     CTLFLAG_RD, SYSCTL_NULL_INT_PTR, TCP_FASTOPEN_KEY_LEN,
 296     "Key length in bytes");
 297
 298 SYSCTL_INT(_net_inet_tcp_fastopen, OID_AUTO, maxkeys,
 299     CTLFLAG_RD, SYSCTL_NULL_INT_PTR, TCP_FASTOPEN_MAX_KEYS,
 300     "Maximum number of keys supported");
 301
 302 SYSCTL_INT(_net_inet_tcp_fastopen, OID_AUTO, maxpsks,
 303     CTLFLAG_RD, SYSCTL_NULL_INT_PTR, TCP_FASTOPEN_MAX_PSKS,
 304     "Maximum number of pre-shared keys supported");
 305
 306 VNET_DEFINE_STATIC(unsigned int, tcp_fastopen_numkeys) = 0;
 307 #define V_tcp_fastopen_numkeys  VNET(tcp_fastopen_numkeys)
 308 SYSCTL_UINT(_net_inet_tcp_fastopen, OID_AUTO, numkeys,
 309     CTLFLAG_VNET | CTLFLAG_RD, &VNET_NAME(tcp_fastopen_numkeys), 0,
 310     "Number of keys installed");
 311
 312 VNET_DEFINE_STATIC(unsigned int, tcp_fastopen_numpsks) = 0;
 313 #define V_tcp_fastopen_numpsks  VNET(tcp_fastopen_numpsks)
 314 SYSCTL_UINT(_net_inet_tcp_fastopen, OID_AUTO, numpsks,
 315     CTLFLAG_VNET | CTLFLAG_RD, &VNET_NAME(tcp_fastopen_numpsks), 0,
 316     "Number of pre-shared keys installed");
 317
 318 VNET_DEFINE_STATIC(unsigned int, tcp_fastopen_path_disable_time) =
 319     TCP_FASTOPEN_PATH_DISABLE_TIME_DEFAULT;
 320 #define V_tcp_fastopen_path_disable_time VNET(tcp_fastopen_path_disable_time)
 321 SYSCTL_UINT(_net_inet_tcp_fastopen, OID_AUTO, path_disable_time,
 322     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_fastopen_path_disable_time), 0,
 323     "Seconds a TFO failure disables a {client_ip, server_ip, server_port} path");
 324
 325 VNET_DEFINE_STATIC(unsigned int, tcp_fastopen_psk_enable) = 0;
 326 #define V_tcp_fastopen_psk_enable       VNET(tcp_fastopen_psk_enable)
 327 static int sysctl_net_inet_tcp_fastopen_psk_enable(SYSCTL_HANDLER_ARGS);
 328 SYSCTL_PROC(_net_inet_tcp_fastopen, OID_AUTO, psk_enable,
 329     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
 330     NULL, 0, &sysctl_net_inet_tcp_fastopen_psk_enable, "IU",
 331     "Enable/disable TCP Fast Open server pre-shared key mode");
 332
 333 VNET_DEFINE(unsigned int, tcp_fastopen_server_enable) = 0;
 334 static int sysctl_net_inet_tcp_fastopen_server_enable(SYSCTL_HANDLER_ARGS);
 335 SYSCTL_PROC(_net_inet_tcp_fastopen, OID_AUTO, server_enable,
 336     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
 337     NULL, 0, &sysctl_net_inet_tcp_fastopen_server_enable, "IU",
 338     "Enable/disable TCP Fast Open server functionality");
 339
 340 static int sysctl_net_inet_tcp_fastopen_setkey(SYSCTL_HANDLER_ARGS);
 341 SYSCTL_PROC(_net_inet_tcp_fastopen, OID_AUTO, setkey,
 342     CTLFLAG_VNET | CTLTYPE_OPAQUE | CTLFLAG_WR | CTLFLAG_MPSAFE,
 343     NULL, 0, &sysctl_net_inet_tcp_fastopen_setkey, "",
 344     "Install a new key");
 345
 346 static int sysctl_net_inet_tcp_fastopen_setpsk(SYSCTL_HANDLER_ARGS);
 347 SYSCTL_PROC(_net_inet_tcp_fastopen, OID_AUTO, setpsk,
 348     CTLFLAG_VNET | CTLTYPE_OPAQUE | CTLFLAG_WR | CTLFLAG_MPSAFE,
 349     NULL, 0, &sysctl_net_inet_tcp_fastopen_setpsk, "",
 350     "Install a new pre-shared key");
 351
 352 static int sysctl_net_inet_tcp_fastopen_ccache_list(SYSCTL_HANDLER_ARGS);
 353 SYSCTL_PROC(_net_inet_tcp_fastopen, OID_AUTO, ccache_list,
 354     CTLFLAG_VNET | CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_MPSAFE,
 355     NULL, 0, sysctl_net_inet_tcp_fastopen_ccache_list, "A",
 356     "List of all client cookie cache entries");
 357
 358 VNET_DEFINE_STATIC(struct rmlock, tcp_fastopen_keylock);
 359 #define V_tcp_fastopen_keylock  VNET(tcp_fastopen_keylock)
 360
 361 #define TCP_FASTOPEN_KEYS_RLOCK(t)      rm_rlock(&V_tcp_fastopen_keylock, (t))
 362 #define TCP_FASTOPEN_KEYS_RUNLOCK(t)    rm_runlock(&V_tcp_fastopen_keylock, (t))
 363 #define TCP_FASTOPEN_KEYS_WLOCK()       rm_wlock(&V_tcp_fastopen_keylock)
 364 #define TCP_FASTOPEN_KEYS_WUNLOCK()     rm_wunlock(&V_tcp_fastopen_keylock)
 365
 366 VNET_DEFINE_STATIC(struct tcp_fastopen_keylist, tcp_fastopen_keys);
 367 #define V_tcp_fastopen_keys     VNET(tcp_fastopen_keys)
 368
 369 VNET_DEFINE_STATIC(struct tcp_fastopen_callout, tcp_fastopen_autokey_ctx);
 370 #define V_tcp_fastopen_autokey_ctx      VNET(tcp_fastopen_autokey_ctx)
 371
 372 VNET_DEFINE_STATIC(uma_zone_t, counter_zone);
 373 #define V_counter_zone                  VNET(counter_zone)
 374
 375 static MALLOC_DEFINE(M_TCP_FASTOPEN_CCACHE, "tfo_ccache", "TFO client cookie cache buckets");
 376
 377 VNET_DEFINE_STATIC(struct tcp_fastopen_ccache, tcp_fastopen_ccache);
 378 #define V_tcp_fastopen_ccache   VNET(tcp_fastopen_ccache)
 379
 380 #define CCB_LOCK(ccb)           mtx_lock(&(ccb)->ccb_mtx)
 381 #define CCB_UNLOCK(ccb)         mtx_unlock(&(ccb)->ccb_mtx)
 382 #define CCB_LOCK_ASSERT(ccb)    mtx_assert(&(ccb)->ccb_mtx, MA_OWNED)
 383
 384 void
 385 tcp_fastopen_init(void)
 386 {
 387         unsigned int i;
 388
 389         V_counter_zone = uma_zcreate("tfo", sizeof(unsigned int),
 390             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
 391         rm_init(&V_tcp_fastopen_keylock, "tfo_keylock");
 392         callout_init_rm(&V_tcp_fastopen_autokey_ctx.c,
 393             &V_tcp_fastopen_keylock, 0);
 394         V_tcp_fastopen_autokey_ctx.v = curvnet;
 395         V_tcp_fastopen_keys.newest = TCP_FASTOPEN_MAX_KEYS - 1;
 396         V_tcp_fastopen_keys.newest_psk = TCP_FASTOPEN_MAX_PSKS - 1;
 397
 398         /* May already be non-zero if kernel tunable was set */
 399         if (V_tcp_fastopen_ccache.bucket_limit == 0)
 400                 V_tcp_fastopen_ccache.bucket_limit =
 401                     TCP_FASTOPEN_CCACHE_BUCKET_LIMIT_DEFAULT;
 402
 403         /* May already be non-zero if kernel tunable was set */
 404         if ((V_tcp_fastopen_ccache_buckets == 0) ||
 405             !powerof2(V_tcp_fastopen_ccache_buckets))
 406                 V_tcp_fastopen_ccache.buckets =
 407                         TCP_FASTOPEN_CCACHE_BUCKETS_DEFAULT;
 408         else
 409                 V_tcp_fastopen_ccache.buckets = V_tcp_fastopen_ccache_buckets;
 410
 411         V_tcp_fastopen_ccache.mask = V_tcp_fastopen_ccache.buckets - 1;
 412         V_tcp_fastopen_ccache.secret = arc4random();
 413
 414         V_tcp_fastopen_ccache.base = malloc(V_tcp_fastopen_ccache.buckets *
 415             sizeof(struct tcp_fastopen_ccache_bucket), M_TCP_FASTOPEN_CCACHE,
 416             M_WAITOK | M_ZERO);
 417
 418         for (i = 0; i < V_tcp_fastopen_ccache.buckets; i++) {
 419                 TAILQ_INIT(&V_tcp_fastopen_ccache.base[i].ccb_entries);
 420                 mtx_init(&V_tcp_fastopen_ccache.base[i].ccb_mtx, "tfo_ccache_bucket",
 421                          NULL, MTX_DEF);
 422                 if (V_tcp_fastopen_client_enable) {
 423                         /* enable bucket */
 424                         V_tcp_fastopen_ccache.base[i].ccb_num_entries = 0;
 425                 } else {
 426                         /* disable bucket */
 427                         V_tcp_fastopen_ccache.base[i].ccb_num_entries = -1;
 428                 }
 429                 V_tcp_fastopen_ccache.base[i].ccb_ccache = &V_tcp_fastopen_ccache;
 430         }
 431
 432         /*
 433          * Note that while the total number of entries in the cookie cache
 434          * is limited by the table management logic to
 435          * V_tcp_fastopen_ccache.buckets *
 436          * V_tcp_fastopen_ccache.bucket_limit, the total number of items in
 437          * this zone can exceed that amount by the number of CPUs in the
 438          * system times the maximum number of unallocated items that can be
 439          * present in each UMA per-CPU cache for this zone.
 440          */
 441         V_tcp_fastopen_ccache.zone = uma_zcreate("tfo_ccache_entries",
 442             sizeof(struct tcp_fastopen_ccache_entry), NULL, NULL, NULL, NULL,
 443             UMA_ALIGN_CACHE, 0);
 444 }
 445
 446 void
 447 tcp_fastopen_destroy(void)
 448 {
 449         struct tcp_fastopen_ccache_bucket *ccb;
 450         unsigned int i;
 451
 452         for (i = 0; i < V_tcp_fastopen_ccache.buckets; i++) {
 453                 ccb = &V_tcp_fastopen_ccache.base[i];
 454                 tcp_fastopen_ccache_bucket_trim(ccb, 0);
 455                 mtx_destroy(&ccb->ccb_mtx);
 456         }
 457
 458         KASSERT(uma_zone_get_cur(V_tcp_fastopen_ccache.zone) == 0,
 459             ("%s: TFO ccache zone allocation count not 0", __func__));
 460         uma_zdestroy(V_tcp_fastopen_ccache.zone);
 461         free(V_tcp_fastopen_ccache.base, M_TCP_FASTOPEN_CCACHE);
 462
 463         callout_drain(&V_tcp_fastopen_autokey_ctx.c);
 464         rm_destroy(&V_tcp_fastopen_keylock);
 465         uma_zdestroy(V_counter_zone);
 466 }
 467
 468 unsigned int *
 469 tcp_fastopen_alloc_counter(void)
 470 {
 471         unsigned int *counter;
 472         counter = uma_zalloc(V_counter_zone, M_NOWAIT);
 473         if (counter)
 474                 *counter = 1;
 475         return (counter);
 476 }
 477
 478 void
 479 tcp_fastopen_decrement_counter(unsigned int *counter)
 480 {
 481         if (*counter == 1)
 482                 uma_zfree(V_counter_zone, counter);
 483         else
 484                 atomic_subtract_int(counter, 1);
 485 }
 486
 487 static void
 488 tcp_fastopen_addkey_locked(uint8_t *key)
 489 {
 490
 491         V_tcp_fastopen_keys.newest++;
 492         if (V_tcp_fastopen_keys.newest == TCP_FASTOPEN_MAX_KEYS)
 493                 V_tcp_fastopen_keys.newest = 0;
 494         memcpy(V_tcp_fastopen_keys.key[V_tcp_fastopen_keys.newest], key,
 495             TCP_FASTOPEN_KEY_LEN);
 496         if (V_tcp_fastopen_numkeys < TCP_FASTOPEN_MAX_KEYS)
 497                 V_tcp_fastopen_numkeys++;
 498 }
 499
 500 static void
 501 tcp_fastopen_addpsk_locked(uint8_t *psk)
 502 {
 503
 504         V_tcp_fastopen_keys.newest_psk++;
 505         if (V_tcp_fastopen_keys.newest_psk == TCP_FASTOPEN_MAX_PSKS)
 506                 V_tcp_fastopen_keys.newest_psk = 0;
 507         memcpy(V_tcp_fastopen_keys.psk[V_tcp_fastopen_keys.newest_psk], psk,
 508             TCP_FASTOPEN_KEY_LEN);
 509         if (V_tcp_fastopen_numpsks < TCP_FASTOPEN_MAX_PSKS)
 510                 V_tcp_fastopen_numpsks++;
 511 }
 512
 513 static void
 514 tcp_fastopen_autokey_locked(void)
 515 {
 516         uint8_t newkey[TCP_FASTOPEN_KEY_LEN];
 517
 518         arc4rand(newkey, TCP_FASTOPEN_KEY_LEN, 0);
 519         tcp_fastopen_addkey_locked(newkey);
 520 }
 521
 522 static void
 523 tcp_fastopen_autokey_callout(void *arg)
 524 {
 525         struct tcp_fastopen_callout *ctx = arg;
 526
 527         CURVNET_SET(ctx->v);
 528         tcp_fastopen_autokey_locked();
 529         callout_reset(&ctx->c, V_tcp_fastopen_autokey * hz,
 530                       tcp_fastopen_autokey_callout, ctx);
 531         CURVNET_RESTORE();
 532 }
 533
 534 static uint64_t
 535 tcp_fastopen_make_cookie(uint8_t key[SIPHASH_KEY_LENGTH], struct in_conninfo *inc)
 536 {
 537         SIPHASH_CTX ctx;
 538         uint64_t siphash;
 539
 540         SipHash24_Init(&ctx);
 541         SipHash_SetKey(&ctx, key);
 542         switch (inc->inc_flags & INC_ISIPV6) {
 543 #ifdef INET
 544         case 0:
 545                 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
 546                 break;
 547 #endif
 548 #ifdef INET6
 549         case INC_ISIPV6:
 550                 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
 551                 break;
 552 #endif
 553         }
 554         SipHash_Final((u_int8_t *)&siphash, &ctx);
 555
 556         return (siphash);
 557 }
 558
 559 static uint64_t
 560 tcp_fastopen_make_psk_cookie(uint8_t *psk, uint8_t *cookie, uint8_t cookie_len)
 561 {
 562         SIPHASH_CTX ctx;
 563         uint64_t psk_cookie;
 564
 565         SipHash24_Init(&ctx);
 566         SipHash_SetKey(&ctx, psk);
 567         SipHash_Update(&ctx, cookie, cookie_len);
 568         SipHash_Final((u_int8_t *)&psk_cookie, &ctx);
 569
 570         return (psk_cookie);
 571 }
 572
 573 static int
 574 tcp_fastopen_find_cookie_match_locked(uint8_t *wire_cookie, uint64_t *cur_cookie)
 575 {
 576         unsigned int i, psk_index;
 577         uint64_t psk_cookie;
 578
 579         if (V_tcp_fastopen_psk_enable) {
 580                 psk_index = V_tcp_fastopen_keys.newest_psk;
 581                 for (i = 0; i < V_tcp_fastopen_numpsks; i++) {
 582                         psk_cookie =
 583                             tcp_fastopen_make_psk_cookie(
 584                                  V_tcp_fastopen_keys.psk[psk_index],
 585                                  (uint8_t *)cur_cookie,
 586                                  TCP_FASTOPEN_COOKIE_LEN);
 587
 588                         if (memcmp(wire_cookie, &psk_cookie,
 589                                    TCP_FASTOPEN_COOKIE_LEN) == 0)
 590                                 return (1);
 591
 592                         if (psk_index == 0)
 593                                 psk_index = TCP_FASTOPEN_MAX_PSKS - 1;
 594                         else
 595                                 psk_index--;
 596                 }
 597         } else if (memcmp(wire_cookie, cur_cookie, TCP_FASTOPEN_COOKIE_LEN) == 0)
 598                 return (1);
 599
 600         return (0);
 601 }
 602
 603 /*
 604  * Return values:
 605  *      -1      the cookie is invalid and no valid cookie is available
 606  *       0      the cookie is invalid and the latest cookie has been returned
 607  *       1      the cookie is valid and the latest cookie has been returned
 608  */
 609 int
 610 tcp_fastopen_check_cookie(struct in_conninfo *inc, uint8_t *cookie,
 611     unsigned int len, uint64_t *latest_cookie)
 612 {
 613         struct rm_priotracker tracker;
 614         unsigned int i, key_index;
 615         int rv;
 616         uint64_t cur_cookie;
 617
 618         if (V_tcp_fastopen_acceptany) {
 619                 *latest_cookie = 0;
 620                 return (1);
 621         }
 622
 623         TCP_FASTOPEN_KEYS_RLOCK(&tracker);
 624         if (len != TCP_FASTOPEN_COOKIE_LEN) {
 625                 if (V_tcp_fastopen_numkeys > 0) {
 626                         *latest_cookie =
 627                             tcp_fastopen_make_cookie(
 628                                 V_tcp_fastopen_keys.key[V_tcp_fastopen_keys.newest],
 629                                 inc);
 630                         rv = 0;
 631                 } else
 632                         rv = -1;
 633                 goto out;
 634         }
 635
 636         /*
 637          * Check against each available key, from newest to oldest.
 638          */
 639         key_index = V_tcp_fastopen_keys.newest;
 640         for (i = 0; i < V_tcp_fastopen_numkeys; i++) {
 641                 cur_cookie =
 642                     tcp_fastopen_make_cookie(V_tcp_fastopen_keys.key[key_index],
 643                         inc);
 644                 if (i == 0)
 645                         *latest_cookie = cur_cookie;
 646                 rv = tcp_fastopen_find_cookie_match_locked(cookie, &cur_cookie);
 647                 if (rv)
 648                         goto out;
 649                 if (key_index == 0)
 650                         key_index = TCP_FASTOPEN_MAX_KEYS - 1;
 651                 else
 652                         key_index--;
 653         }
 654         rv = 0;
 655
 656  out:
 657         TCP_FASTOPEN_KEYS_RUNLOCK(&tracker);
 658         return (rv);
 659 }
 660
 661 static int
 662 sysctl_net_inet_tcp_fastopen_autokey(SYSCTL_HANDLER_ARGS)
 663 {
 664         int error;
 665         unsigned int new;
 666
 667         new = V_tcp_fastopen_autokey;
 668         error = sysctl_handle_int(oidp, &new, 0, req);
 669         if (error == 0 && req->newptr) {
 670                 if (new > (INT_MAX / hz))
 671                         return (EINVAL);
 672
 673                 TCP_FASTOPEN_KEYS_WLOCK();
 674                 if (V_tcp_fastopen_server_enable) {
 675                         if (V_tcp_fastopen_autokey && !new)
 676                                 callout_stop(&V_tcp_fastopen_autokey_ctx.c);
 677                         else if (new)
 678                                 callout_reset(&V_tcp_fastopen_autokey_ctx.c,
 679                                     new * hz, tcp_fastopen_autokey_callout,
 680                                     &V_tcp_fastopen_autokey_ctx);
 681                 }
 682                 V_tcp_fastopen_autokey = new;
 683                 TCP_FASTOPEN_KEYS_WUNLOCK();
 684         }
 685
 686         return (error);
 687 }
 688
 689 static int
 690 sysctl_net_inet_tcp_fastopen_psk_enable(SYSCTL_HANDLER_ARGS)
 691 {
 692         int error;
 693         unsigned int new;
 694
 695         new = V_tcp_fastopen_psk_enable;
 696         error = sysctl_handle_int(oidp, &new, 0, req);
 697         if (error == 0 && req->newptr) {
 698                 if (V_tcp_fastopen_psk_enable && !new) {
 699                         /* enabled -> disabled */
 700                         TCP_FASTOPEN_KEYS_WLOCK();
 701                         V_tcp_fastopen_numpsks = 0;
 702                         V_tcp_fastopen_keys.newest_psk =
 703                             TCP_FASTOPEN_MAX_PSKS - 1;
 704                         V_tcp_fastopen_psk_enable = 0;
 705                         TCP_FASTOPEN_KEYS_WUNLOCK();
 706                 } else if (!V_tcp_fastopen_psk_enable && new) {
 707                         /* disabled -> enabled */
 708                         TCP_FASTOPEN_KEYS_WLOCK();
 709                         V_tcp_fastopen_psk_enable = 1;
 710                         TCP_FASTOPEN_KEYS_WUNLOCK();
 711                 }
 712         }
 713         return (error);
 714 }
 715
 716 static int
 717 sysctl_net_inet_tcp_fastopen_server_enable(SYSCTL_HANDLER_ARGS)
 718 {
 719         int error;
 720         unsigned int new;
 721
 722         new = V_tcp_fastopen_server_enable;
 723         error = sysctl_handle_int(oidp, &new, 0, req);
 724         if (error == 0 && req->newptr) {
 725                 if (V_tcp_fastopen_server_enable && !new) {
 726                         /* enabled -> disabled */
 727                         TCP_FASTOPEN_KEYS_WLOCK();
 728                         V_tcp_fastopen_numkeys = 0;
 729                         V_tcp_fastopen_keys.newest = TCP_FASTOPEN_MAX_KEYS - 1;
 730                         if (V_tcp_fastopen_autokey)
 731                                 callout_stop(&V_tcp_fastopen_autokey_ctx.c);
 732                         V_tcp_fastopen_numpsks = 0;
 733                         V_tcp_fastopen_keys.newest_psk =
 734                             TCP_FASTOPEN_MAX_PSKS - 1;
 735                         V_tcp_fastopen_server_enable = 0;
 736                         TCP_FASTOPEN_KEYS_WUNLOCK();
 737                 } else if (!V_tcp_fastopen_server_enable && new) {
 738                         /* disabled -> enabled */
 739                         TCP_FASTOPEN_KEYS_WLOCK();
 740                         if (V_tcp_fastopen_autokey &&
 741                             (V_tcp_fastopen_numkeys == 0)) {
 742                                 tcp_fastopen_autokey_locked();
 743                                 callout_reset(&V_tcp_fastopen_autokey_ctx.c,
 744                                     V_tcp_fastopen_autokey * hz,
 745                                     tcp_fastopen_autokey_callout,
 746                                     &V_tcp_fastopen_autokey_ctx);
 747                         }
 748                         V_tcp_fastopen_server_enable = 1;
 749                         TCP_FASTOPEN_KEYS_WUNLOCK();
 750                 }
 751         }
 752         return (error);
 753 }
 754
 755 static int
 756 sysctl_net_inet_tcp_fastopen_setkey(SYSCTL_HANDLER_ARGS)
 757 {
 758         int error;
 759         uint8_t newkey[TCP_FASTOPEN_KEY_LEN];
 760
 761         if (req->oldptr != NULL || req->oldlen != 0)
 762                 return (EINVAL);
 763         if (req->newptr == NULL)
 764                 return (EPERM);
 765         if (req->newlen != sizeof(newkey))
 766                 return (EINVAL);
 767         error = SYSCTL_IN(req, newkey, sizeof(newkey));
 768         if (error)
 769                 return (error);
 770
 771         TCP_FASTOPEN_KEYS_WLOCK();
 772         tcp_fastopen_addkey_locked(newkey);
 773         TCP_FASTOPEN_KEYS_WUNLOCK();
 774
 775         return (0);
 776 }
 777
 778 static int
 779 sysctl_net_inet_tcp_fastopen_setpsk(SYSCTL_HANDLER_ARGS)
 780 {
 781         int error;
 782         uint8_t newpsk[TCP_FASTOPEN_KEY_LEN];
 783
 784         if (req->oldptr != NULL || req->oldlen != 0)
 785                 return (EINVAL);
 786         if (req->newptr == NULL)
 787                 return (EPERM);
 788         if (req->newlen != sizeof(newpsk))
 789                 return (EINVAL);
 790         error = SYSCTL_IN(req, newpsk, sizeof(newpsk));
 791         if (error)
 792                 return (error);
 793
 794         TCP_FASTOPEN_KEYS_WLOCK();
 795         tcp_fastopen_addpsk_locked(newpsk);
 796         TCP_FASTOPEN_KEYS_WUNLOCK();
 797
 798         return (0);
 799 }
 800
 801 static int
 802 sysctl_net_inet_tcp_fastopen_ccache_bucket_limit(SYSCTL_HANDLER_ARGS)
 803 {
 804         struct tcp_fastopen_ccache_bucket *ccb;
 805         int error;
 806         unsigned int new;
 807         unsigned int i;
 808
 809         new = V_tcp_fastopen_ccache.bucket_limit;
 810         error = sysctl_handle_int(oidp, &new, 0, req);
 811         if (error == 0 && req->newptr) {
 812                 if ((new == 0) || (new > INT_MAX))
 813                         error = EINVAL;
 814                 else {
 815                         if (new < V_tcp_fastopen_ccache.bucket_limit) {
 816                                 for (i = 0; i < V_tcp_fastopen_ccache.buckets;
 817                                      i++) {
 818                                         ccb = &V_tcp_fastopen_ccache.base[i];
 819                                         tcp_fastopen_ccache_bucket_trim(ccb, new);
 820                                 }
 821                         }
 822                         V_tcp_fastopen_ccache.bucket_limit = new;
 823                 }
 824         }
 825         return (error);
 826 }
 827
 828 static int
 829 sysctl_net_inet_tcp_fastopen_client_enable(SYSCTL_HANDLER_ARGS)
 830 {
 831         struct tcp_fastopen_ccache_bucket *ccb;
 832         int error;
 833         unsigned int new, i;
 834
 835         new = V_tcp_fastopen_client_enable;
 836         error = sysctl_handle_int(oidp, &new, 0, req);
 837         if (error == 0 && req->newptr) {
 838                 if (V_tcp_fastopen_client_enable && !new) {
 839                         /* enabled -> disabled */
 840                         for (i = 0; i < V_tcp_fastopen_ccache.buckets; i++) {
 841                                 ccb = &V_tcp_fastopen_ccache.base[i];
 842                                 KASSERT(ccb->ccb_num_entries > -1,
 843                                     ("%s: ccb->ccb_num_entries %d is negative",
 844                                         __func__, ccb->ccb_num_entries));
 845                                 tcp_fastopen_ccache_bucket_trim(ccb, 0);
 846                         }
 847                         V_tcp_fastopen_client_enable = 0;
 848                 } else if (!V_tcp_fastopen_client_enable && new) {
 849                         /* disabled -> enabled */
 850                         for (i = 0; i < V_tcp_fastopen_ccache.buckets; i++) {
 851                                 ccb = &V_tcp_fastopen_ccache.base[i];
 852                                 CCB_LOCK(ccb);
 853                                 KASSERT(TAILQ_EMPTY(&ccb->ccb_entries),
 854                                     ("%s: ccb->ccb_entries not empty", __func__));
 855                                 KASSERT(ccb->ccb_num_entries == -1,
 856                                     ("%s: ccb->ccb_num_entries %d not -1", __func__,
 857                                         ccb->ccb_num_entries));
 858                                 ccb->ccb_num_entries = 0; /* enable bucket */
 859                                 CCB_UNLOCK(ccb);
 860                         }
 861                         V_tcp_fastopen_client_enable = 1;
 862                 }
 863         }
 864         return (error);
 865 }
 866
 867 void
 868 tcp_fastopen_connect(struct tcpcb *tp)
 869 {
 870         struct inpcb *inp;
 871         struct tcp_fastopen_ccache_bucket *ccb;
 872         struct tcp_fastopen_ccache_entry *cce;
 873         sbintime_t now;
 874         uint16_t server_mss;
 875         uint64_t psk_cookie;
 876
 877         psk_cookie = 0;
 878         inp = tp->t_inpcb;
 879         cce = tcp_fastopen_ccache_lookup(&inp->inp_inc, &ccb);
 880         if (cce) {
 881                 if (cce->disable_time == 0) {
 882                         if ((cce->cookie_len > 0) &&
 883                             (tp->t_tfo_client_cookie_len ==
 884                              TCP_FASTOPEN_PSK_LEN)) {
 885                                 psk_cookie =
 886                                     tcp_fastopen_make_psk_cookie(
 887                                         tp->t_tfo_cookie.client,
 888                                         cce->cookie, cce->cookie_len);
 889                         } else {
 890                                 tp->t_tfo_client_cookie_len = cce->cookie_len;
 891                                 memcpy(tp->t_tfo_cookie.client, cce->cookie,
 892                                     cce->cookie_len);
 893                         }
 894                         server_mss = cce->server_mss;
 895                         CCB_UNLOCK(ccb);
 896                         if (tp->t_tfo_client_cookie_len ==
 897                             TCP_FASTOPEN_PSK_LEN && psk_cookie) {
 898                                 tp->t_tfo_client_cookie_len =
 899                                     TCP_FASTOPEN_COOKIE_LEN;
 900                                 memcpy(tp->t_tfo_cookie.client, &psk_cookie,
 901                                     TCP_FASTOPEN_COOKIE_LEN);
 902                         }
 903                         tcp_mss(tp, server_mss ? server_mss : -1);
 904                         tp->snd_wnd = tp->t_maxseg;
 905                 } else {
 906                         /*
 907                          * The path is disabled.  Check the time and
 908                          * possibly re-enable.
 909                          */
 910                         now = getsbinuptime();
 911                         if (now - cce->disable_time >
 912                             ((sbintime_t)V_tcp_fastopen_path_disable_time << 32)) {
 913                                 /*
 914                                  * Re-enable path.  Force a TFO cookie
 915                                  * request.  Forget the old MSS as it may be
 916                                  * bogus now, and we will rediscover it in
 917                                  * the SYN|ACK.
 918                                  */
 919                                 cce->disable_time = 0;
 920                                 cce->server_mss = 0;
 921                                 cce->cookie_len = 0;
 922                                 /*
 923                                  * tp->t_tfo... cookie details are already
 924                                  * zero from the tcpcb init.
 925                                  */
 926                         } else {
 927                                 /*
 928                                  * Path is disabled, so disable TFO on this
 929                                  * connection.
 930                                  */
 931                                 tp->t_flags &= ~TF_FASTOPEN;
 932                         }
 933                         CCB_UNLOCK(ccb);
 934                         tcp_mss(tp, -1);
 935                         /*
 936                          * snd_wnd is irrelevant since we are either forcing
 937                          * a TFO cookie request or disabling TFO - either
 938                          * way, no data with the SYN.
 939                          */
 940                 }
 941         } else {
 942                 /*
 943                  * A new entry for this path will be created when a SYN|ACK
 944                  * comes back, or the attempt otherwise fails.
 945                  */
 946                 CCB_UNLOCK(ccb);
 947                 tcp_mss(tp, -1);
 948                 /*
 949                  * snd_wnd is irrelevant since we are forcing a TFO cookie
 950                  * request.
 951                  */
 952         }
 953 }
 954
 955 void
 956 tcp_fastopen_disable_path(struct tcpcb *tp)
 957 {
 958         struct in_conninfo *inc = &tp->t_inpcb->inp_inc;
 959         struct tcp_fastopen_ccache_bucket *ccb;
 960         struct tcp_fastopen_ccache_entry *cce;
 961
 962         cce = tcp_fastopen_ccache_lookup(inc, &ccb);
 963         if (cce) {
 964                 cce->server_mss = 0;
 965                 cce->cookie_len = 0;
 966                 /*
 967                  * Preserve the existing disable time if it is already
 968                  * disabled.
 969                  */
 970                 if (cce->disable_time == 0)
 971                         cce->disable_time = getsbinuptime();
 972         } else /* use invalid cookie len to create disabled entry */
 973                 tcp_fastopen_ccache_create(ccb, inc, 0,
 974                     TCP_FASTOPEN_MAX_COOKIE_LEN + 1, NULL);
 975
 976         CCB_UNLOCK(ccb);
 977         tp->t_flags &= ~TF_FASTOPEN;
 978 }
 979
 980 void
 981 tcp_fastopen_update_cache(struct tcpcb *tp, uint16_t mss,
 982     uint8_t cookie_len, uint8_t *cookie)
 983 {
 984         struct in_conninfo *inc = &tp->t_inpcb->inp_inc;
 985         struct tcp_fastopen_ccache_bucket *ccb;
 986         struct tcp_fastopen_ccache_entry *cce;
 987
 988         cce = tcp_fastopen_ccache_lookup(inc, &ccb);
 989         if (cce) {
 990                 if ((cookie_len >= TCP_FASTOPEN_MIN_COOKIE_LEN) &&
 991                     (cookie_len <= TCP_FASTOPEN_MAX_COOKIE_LEN) &&
 992                     ((cookie_len & 0x1) == 0)) {
 993                         cce->server_mss = mss;
 994                         cce->cookie_len = cookie_len;
 995                         memcpy(cce->cookie, cookie, cookie_len);
 996                         cce->disable_time = 0;
 997                 } else {
 998                         /* invalid cookie length, disable entry */
 999                         cce->server_mss = 0;
1000                         cce->cookie_len = 0;
1001                         /*
1002                          * Preserve the existing disable time if it is
1003                          * already disabled.
1004                          */
1005                         if (cce->disable_time == 0)
1006                                 cce->disable_time = getsbinuptime();
1007                 }
1008         } else
1009                 tcp_fastopen_ccache_create(ccb, inc, mss, cookie_len, cookie);
1010
1011         CCB_UNLOCK(ccb);
1012 }
1013
1014 static struct tcp_fastopen_ccache_entry *
1015 tcp_fastopen_ccache_lookup(struct in_conninfo *inc,
1016     struct tcp_fastopen_ccache_bucket **ccbp)
1017 {
1018         struct tcp_fastopen_ccache_bucket *ccb;
1019         struct tcp_fastopen_ccache_entry *cce;
1020         uint32_t last_word;
1021         uint32_t hash;
1022
1023         hash = jenkins_hash32((uint32_t *)&inc->inc_ie.ie_dependladdr, 4,
1024             V_tcp_fastopen_ccache.secret);
1025         hash = jenkins_hash32((uint32_t *)&inc->inc_ie.ie_dependfaddr, 4,
1026             hash);
1027         last_word = inc->inc_fport;
1028         hash = jenkins_hash32(&last_word, 1, hash);
1029         ccb = &V_tcp_fastopen_ccache.base[hash & V_tcp_fastopen_ccache.mask];
1030         *ccbp = ccb;
1031         CCB_LOCK(ccb);
1032
1033         /*
1034          * Always returns with locked bucket.
1035          */
1036         TAILQ_FOREACH(cce, &ccb->ccb_entries, cce_link)
1037                 if ((!(cce->af == AF_INET6) == !(inc->inc_flags & INC_ISIPV6)) &&
1038                     (cce->server_port == inc->inc_ie.ie_fport) &&
1039                     (((cce->af == AF_INET) &&
1040                       (cce->cce_client_ip.v4.s_addr == inc->inc_laddr.s_addr) &&
1041                       (cce->cce_server_ip.v4.s_addr == inc->inc_faddr.s_addr)) ||
1042                      ((cce->af == AF_INET6) &&
1043                       IN6_ARE_ADDR_EQUAL(&cce->cce_client_ip.v6, &inc->inc6_laddr) &&
1044                       IN6_ARE_ADDR_EQUAL(&cce->cce_server_ip.v6, &inc->inc6_faddr))))
1045                         break;
1046
1047         return (cce);
1048 }
1049
1050 static struct tcp_fastopen_ccache_entry *
1051 tcp_fastopen_ccache_create(struct tcp_fastopen_ccache_bucket *ccb,
1052     struct in_conninfo *inc, uint16_t mss, uint8_t cookie_len, uint8_t *cookie)
1053 {
1054         struct tcp_fastopen_ccache_entry *cce;
1055
1056         /*
1057          * 1. Create a new entry, or
1058          * 2. Reclaim an existing entry, or
1059          * 3. Fail
1060          */
1061
1062         CCB_LOCK_ASSERT(ccb);
1063
1064         cce = NULL;
1065         if (ccb->ccb_num_entries < V_tcp_fastopen_ccache.bucket_limit)
1066                 cce = uma_zalloc(V_tcp_fastopen_ccache.zone, M_NOWAIT);
1067
1068         if (cce == NULL) {
1069                 /*
1070                  * At bucket limit, or out of memory - reclaim last
1071                  * entry in bucket.
1072                  */
1073                 cce = TAILQ_LAST(&ccb->ccb_entries, bucket_entries);
1074                 if (cce == NULL) {
1075                         /* XXX count this event */
1076                         return (NULL);
1077                 }
1078
1079                 TAILQ_REMOVE(&ccb->ccb_entries, cce, cce_link);
1080         } else
1081                 ccb->ccb_num_entries++;
1082
1083         TAILQ_INSERT_HEAD(&ccb->ccb_entries, cce, cce_link);
1084         cce->af = (inc->inc_flags & INC_ISIPV6) ? AF_INET6 : AF_INET;
1085         if (cce->af == AF_INET) {
1086                 cce->cce_client_ip.v4 = inc->inc_laddr;
1087                 cce->cce_server_ip.v4 = inc->inc_faddr;
1088         } else {
1089                 cce->cce_client_ip.v6 = inc->inc6_laddr;
1090                 cce->cce_server_ip.v6 = inc->inc6_faddr;
1091         }
1092         cce->server_port = inc->inc_fport;
1093         if ((cookie_len >= TCP_FASTOPEN_MIN_COOKIE_LEN) &&
1094             (cookie_len <= TCP_FASTOPEN_MAX_COOKIE_LEN) &&
1095             ((cookie_len & 0x1) == 0)) {
1096                 cce->server_mss = mss;
1097                 cce->cookie_len = cookie_len;
1098                 memcpy(cce->cookie, cookie, cookie_len);
1099                 cce->disable_time = 0;
1100         } else {
1101                 /* invalid cookie length, disable cce */
1102                 cce->server_mss = 0;
1103                 cce->cookie_len = 0;
1104                 cce->disable_time = getsbinuptime();
1105         }
1106
1107         return (cce);
1108 }
1109
1110 static void
1111 tcp_fastopen_ccache_bucket_trim(struct tcp_fastopen_ccache_bucket *ccb,
1112     unsigned int limit)
1113 {
1114         struct tcp_fastopen_ccache_entry *cce, *cce_tmp;
1115         unsigned int entries;
1116
1117         CCB_LOCK(ccb);
1118         entries = 0;
1119         TAILQ_FOREACH_SAFE(cce, &ccb->ccb_entries, cce_link, cce_tmp) {
1120                 entries++;
1121                 if (entries > limit)
1122                         tcp_fastopen_ccache_entry_drop(cce, ccb);
1123         }
1124         KASSERT(ccb->ccb_num_entries <= (int)limit,
1125             ("%s: ccb->ccb_num_entries %d exceeds limit %d", __func__,
1126                 ccb->ccb_num_entries, limit));
1127         if (limit == 0) {
1128                 KASSERT(TAILQ_EMPTY(&ccb->ccb_entries),
1129                     ("%s: ccb->ccb_entries not empty", __func__));
1130                 ccb->ccb_num_entries = -1; /* disable bucket */
1131         }
1132         CCB_UNLOCK(ccb);
1133 }
1134
1135 static void
1136 tcp_fastopen_ccache_entry_drop(struct tcp_fastopen_ccache_entry *cce,
1137     struct tcp_fastopen_ccache_bucket *ccb)
1138 {
1139
1140         CCB_LOCK_ASSERT(ccb);
1141
1142         TAILQ_REMOVE(&ccb->ccb_entries, cce, cce_link);
1143         ccb->ccb_num_entries--;
1144         uma_zfree(V_tcp_fastopen_ccache.zone, cce);
1145 }
1146
1147 static int
1148 sysctl_net_inet_tcp_fastopen_ccache_list(SYSCTL_HANDLER_ARGS)
1149 {
1150         struct sbuf sb;
1151         struct tcp_fastopen_ccache_bucket *ccb;
1152         struct tcp_fastopen_ccache_entry *cce;
1153         sbintime_t now, duration, limit;
1154         const int linesize = 128;
1155         int i, error, num_entries;
1156         unsigned int j;
1157 #ifdef INET6
1158         char clt_buf[INET6_ADDRSTRLEN], srv_buf[INET6_ADDRSTRLEN];
1159 #else
1160         char clt_buf[INET_ADDRSTRLEN], srv_buf[INET_ADDRSTRLEN];
1161 #endif
1162
1163         if (jailed_without_vnet(curthread->td_ucred) != 0)
1164                 return (EPERM);
1165
1166         /* Only allow root to read the client cookie cache */
1167         if (curthread->td_ucred->cr_uid != 0)
1168                 return (EPERM);
1169
1170         num_entries = 0;
1171         for (i = 0; i < V_tcp_fastopen_ccache.buckets; i++) {
1172                 ccb = &V_tcp_fastopen_ccache.base[i];
1173                 CCB_LOCK(ccb);
1174                 if (ccb->ccb_num_entries > 0)
1175                         num_entries += ccb->ccb_num_entries;
1176                 CCB_UNLOCK(ccb);
1177         }
1178         sbuf_new(&sb, NULL, linesize * (num_entries + 1), SBUF_INCLUDENUL);
1179
1180         sbuf_printf(&sb,
1181                     "\nLocal IP address     Remote IP address     Port   MSS"
1182                     " Disabled Cookie\n");
1183
1184         now = getsbinuptime();
1185         limit = (sbintime_t)V_tcp_fastopen_path_disable_time << 32;
1186         for (i = 0; i < V_tcp_fastopen_ccache.buckets; i++) {
1187                 ccb = &V_tcp_fastopen_ccache.base[i];
1188                 CCB_LOCK(ccb);
1189                 TAILQ_FOREACH(cce, &ccb->ccb_entries, cce_link) {
1190                         if (cce->disable_time != 0) {
1191                                 duration = now - cce->disable_time;
1192                                 if (limit >= duration)
1193                                         duration = limit - duration;
1194                                 else
1195                                         duration = 0;
1196                         } else
1197                                 duration = 0;
1198                         sbuf_printf(&sb,
1199                                     "%-20s %-20s %5u %5u ",
1200                                     inet_ntop(cce->af, &cce->cce_client_ip,
1201                                         clt_buf, sizeof(clt_buf)),
1202                                     inet_ntop(cce->af, &cce->cce_server_ip,
1203                                         srv_buf, sizeof(srv_buf)),
1204                                     ntohs(cce->server_port),
1205                                     cce->server_mss);
1206                         if (duration > 0)
1207                                 sbuf_printf(&sb, "%7ds ", sbintime_getsec(duration));
1208                         else
1209                                 sbuf_printf(&sb, "%8s ", "No");
1210                         for (j = 0; j < cce->cookie_len; j++)
1211                                 sbuf_printf(&sb, "%02x", cce->cookie[j]);
1212                         sbuf_putc(&sb, '\n');
1213                 }
1214                 CCB_UNLOCK(ccb);
1215         }
1216         error = sbuf_finish(&sb);
1217         if (error == 0)
1218                 error = SYSCTL_OUT(req, sbuf_data(&sb), sbuf_len(&sb));
1219         sbuf_delete(&sb);
1220         return (error);
1221 }