net/ipv4/tcp_fastopen.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/crypto.h>
   3 #include <linux/err.h>
   4 #include <linux/init.h>
   5 #include <linux/kernel.h>
   6 #include <linux/list.h>
   7 #include <linux/tcp.h>
   8 #include <linux/rcupdate.h>
   9 #include <linux/rculist.h>
  10 #include <net/inetpeer.h>
  11 #include <net/tcp.h>
  12
  13 void tcp_fastopen_init_key_once(struct net *net)
  14 {
  15         u8 key[TCP_FASTOPEN_KEY_LENGTH];
  16         struct tcp_fastopen_context *ctxt;
  17
  18         rcu_read_lock();
  19         ctxt = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
  20         if (ctxt) {
  21                 rcu_read_unlock();
  22                 return;
  23         }
  24         rcu_read_unlock();
  25
  26         /* tcp_fastopen_reset_cipher publishes the new context
  27          * atomically, so we allow this race happening here.
  28          *
  29          * All call sites of tcp_fastopen_cookie_gen also check
  30          * for a valid cookie, so this is an acceptable risk.
  31          */
  32         get_random_bytes(key, sizeof(key));
  33         tcp_fastopen_reset_cipher(net, NULL, key, sizeof(key));
  34 }
  35
  36 static void tcp_fastopen_ctx_free(struct rcu_head *head)
  37 {
  38         struct tcp_fastopen_context *ctx =
  39             container_of(head, struct tcp_fastopen_context, rcu);
  40         crypto_free_cipher(ctx->tfm);
  41         kfree(ctx);
  42 }
  43
  44 void tcp_fastopen_destroy_cipher(struct sock *sk)
  45 {
  46         struct tcp_fastopen_context *ctx;
  47
  48         ctx = rcu_dereference_protected(
  49                         inet_csk(sk)->icsk_accept_queue.fastopenq.ctx, 1);
  50         if (ctx)
  51                 call_rcu(&ctx->rcu, tcp_fastopen_ctx_free);
  52 }
  53
  54 void tcp_fastopen_ctx_destroy(struct net *net)
  55 {
  56         struct tcp_fastopen_context *ctxt;
  57
  58         spin_lock(&net->ipv4.tcp_fastopen_ctx_lock);
  59
  60         ctxt = rcu_dereference_protected(net->ipv4.tcp_fastopen_ctx,
  61                                 lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
  62         rcu_assign_pointer(net->ipv4.tcp_fastopen_ctx, NULL);
  63         spin_unlock(&net->ipv4.tcp_fastopen_ctx_lock);
  64
  65         if (ctxt)
  66                 call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free);
  67 }
  68
  69 int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
  70                               void *key, unsigned int len)
  71 {
  72         struct tcp_fastopen_context *ctx, *octx;
  73         struct fastopen_queue *q;
  74         int err;
  75
  76         ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
  77         if (!ctx)
  78                 return -ENOMEM;
  79         ctx->tfm = crypto_alloc_cipher("aes", 0, 0);
  80
  81         if (IS_ERR(ctx->tfm)) {
  82                 err = PTR_ERR(ctx->tfm);
  83 error:          kfree(ctx);
  84                 pr_err("TCP: TFO aes cipher alloc error: %d\n", err);
  85                 return err;
  86         }
  87         err = crypto_cipher_setkey(ctx->tfm, key, len);
  88         if (err) {
  89                 pr_err("TCP: TFO cipher key error: %d\n", err);
  90                 crypto_free_cipher(ctx->tfm);
  91                 goto error;
  92         }
  93         memcpy(ctx->key, key, len);
  94
  95
  96         spin_lock(&net->ipv4.tcp_fastopen_ctx_lock);
  97         if (sk) {
  98                 q = &inet_csk(sk)->icsk_accept_queue.fastopenq;
  99                 octx = rcu_dereference_protected(q->ctx,
 100                         lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
 101                 rcu_assign_pointer(q->ctx, ctx);
 102         } else {
 103                 octx = rcu_dereference_protected(net->ipv4.tcp_fastopen_ctx,
 104                         lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
 105                 rcu_assign_pointer(net->ipv4.tcp_fastopen_ctx, ctx);
 106         }
 107         spin_unlock(&net->ipv4.tcp_fastopen_ctx_lock);
 108
 109         if (octx)
 110                 call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
 111         return err;
 112 }
 113
 114 static bool __tcp_fastopen_cookie_gen(struct sock *sk, const void *path,
 115                                       struct tcp_fastopen_cookie *foc)
 116 {
 117         struct tcp_fastopen_context *ctx;
 118         bool ok = false;
 119
 120         rcu_read_lock();
 121
 122         ctx = rcu_dereference(inet_csk(sk)->icsk_accept_queue.fastopenq.ctx);
 123         if (!ctx)
 124                 ctx = rcu_dereference(sock_net(sk)->ipv4.tcp_fastopen_ctx);
 125
 126         if (ctx) {
 127                 crypto_cipher_encrypt_one(ctx->tfm, foc->val, path);
 128                 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
 129                 ok = true;
 130         }
 131         rcu_read_unlock();
 132         return ok;
 133 }
 134
 135 /* Generate the fastopen cookie by doing aes128 encryption on both
 136  * the source and destination addresses. Pad 0s for IPv4 or IPv4-mapped-IPv6
 137  * addresses. For the longer IPv6 addresses use CBC-MAC.
 138  *
 139  * XXX (TFO) - refactor when TCP_FASTOPEN_COOKIE_SIZE != AES_BLOCK_SIZE.
 140  */
 141 static bool tcp_fastopen_cookie_gen(struct sock *sk,
 142                                     struct request_sock *req,
 143                                     struct sk_buff *syn,
 144                                     struct tcp_fastopen_cookie *foc)
 145 {
 146         if (req->rsk_ops->family == AF_INET) {
 147                 const struct iphdr *iph = ip_hdr(syn);
 148
 149                 __be32 path[4] = { iph->saddr, iph->daddr, 0, 0 };
 150                 return __tcp_fastopen_cookie_gen(sk, path, foc);
 151         }
 152
 153 #if IS_ENABLED(CONFIG_IPV6)
 154         if (req->rsk_ops->family == AF_INET6) {
 155                 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
 156                 struct tcp_fastopen_cookie tmp;
 157
 158                 if (__tcp_fastopen_cookie_gen(sk, &ip6h->saddr, &tmp)) {
 159                         struct in6_addr *buf = &tmp.addr;
 160                         int i;
 161
 162                         for (i = 0; i < 4; i++)
 163                                 buf->s6_addr32[i] ^= ip6h->daddr.s6_addr32[i];
 164                         return __tcp_fastopen_cookie_gen(sk, buf, foc);
 165                 }
 166         }
 167 #endif
 168         return false;
 169 }
 170
 171
 172 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
 173  * queue this additional data / FIN.
 174  */
 175 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
 176 {
 177         struct tcp_sock *tp = tcp_sk(sk);
 178
 179         if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
 180                 return;
 181
 182         skb = skb_clone(skb, GFP_ATOMIC);
 183         if (!skb)
 184                 return;
 185
 186         skb_dst_drop(skb);
 187         /* segs_in has been initialized to 1 in tcp_create_openreq_child().
 188          * Hence, reset segs_in to 0 before calling tcp_segs_in()
 189          * to avoid double counting.  Also, tcp_segs_in() expects
 190          * skb->len to include the tcp_hdrlen.  Hence, it should
 191          * be called before __skb_pull().
 192          */
 193         tp->segs_in = 0;
 194         tcp_segs_in(tp, skb);
 195         __skb_pull(skb, tcp_hdrlen(skb));
 196         sk_forced_mem_schedule(sk, skb->truesize);
 197         skb_set_owner_r(skb, sk);
 198
 199         TCP_SKB_CB(skb)->seq++;
 200         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
 201
 202         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
 203         __skb_queue_tail(&sk->sk_receive_queue, skb);
 204         tp->syn_data_acked = 1;
 205
 206         /* u64_stats_update_begin(&tp->syncp) not needed here,
 207          * as we certainly are not changing upper 32bit value (0)
 208          */
 209         tp->bytes_received = skb->len;
 210
 211         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
 212                 tcp_fin(sk);
 213 }
 214
 215 static struct sock *tcp_fastopen_create_child(struct sock *sk,
 216                                               struct sk_buff *skb,
 217                                               struct request_sock *req)
 218 {
 219         struct tcp_sock *tp;
 220         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 221         struct sock *child;
 222         bool own_req;
 223
 224         req->num_retrans = 0;
 225         req->num_timeout = 0;
 226         req->sk = NULL;
 227
 228         child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
 229                                                          NULL, &own_req);
 230         if (!child)
 231                 return NULL;
 232
 233         spin_lock(&queue->fastopenq.lock);
 234         queue->fastopenq.qlen++;
 235         spin_unlock(&queue->fastopenq.lock);
 236
 237         /* Initialize the child socket. Have to fix some values to take
 238          * into account the child is a Fast Open socket and is created
 239          * only out of the bits carried in the SYN packet.
 240          */
 241         tp = tcp_sk(child);
 242
 243         tp->fastopen_rsk = req;
 244         tcp_rsk(req)->tfo_listener = true;
 245
 246         /* RFC1323: The window in SYN & SYN/ACK segments is never
 247          * scaled. So correct it appropriately.
 248          */
 249         tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
 250         tp->max_window = tp->snd_wnd;
 251
 252         /* Activate the retrans timer so that SYNACK can be retransmitted.
 253          * The request socket is not added to the ehash
 254          * because it's been added to the accept queue directly.
 255          */
 256         inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
 257                                   TCP_TIMEOUT_INIT, TCP_RTO_MAX);
 258
 259         refcount_set(&req->rsk_refcnt, 2);
 260
 261         /* Now finish processing the fastopen child socket. */
 262         tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
 263
 264         tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
 265
 266         tcp_fastopen_add_skb(child, skb);
 267
 268         tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
 269         tp->rcv_wup = tp->rcv_nxt;
 270         /* tcp_conn_request() is sending the SYNACK,
 271          * and queues the child into listener accept queue.
 272          */
 273         return child;
 274 }
 275
 276 static bool tcp_fastopen_queue_check(struct sock *sk)
 277 {
 278         struct fastopen_queue *fastopenq;
 279
 280         /* Make sure the listener has enabled fastopen, and we don't
 281          * exceed the max # of pending TFO requests allowed before trying
 282          * to validating the cookie in order to avoid burning CPU cycles
 283          * unnecessarily.
 284          *
 285          * XXX (TFO) - The implication of checking the max_qlen before
 286          * processing a cookie request is that clients can't differentiate
 287          * between qlen overflow causing Fast Open to be disabled
 288          * temporarily vs a server not supporting Fast Open at all.
 289          */
 290         fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
 291         if (fastopenq->max_qlen == 0)
 292                 return false;
 293
 294         if (fastopenq->qlen >= fastopenq->max_qlen) {
 295                 struct request_sock *req1;
 296                 spin_lock(&fastopenq->lock);
 297                 req1 = fastopenq->rskq_rst_head;
 298                 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
 299                         __NET_INC_STATS(sock_net(sk),
 300                                         LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
 301                         spin_unlock(&fastopenq->lock);
 302                         return false;
 303                 }
 304                 fastopenq->rskq_rst_head = req1->dl_next;
 305                 fastopenq->qlen--;
 306                 spin_unlock(&fastopenq->lock);
 307                 reqsk_put(req1);
 308         }
 309         return true;
 310 }
 311
 312 static bool tcp_fastopen_no_cookie(const struct sock *sk,
 313                                    const struct dst_entry *dst,
 314                                    int flag)
 315 {
 316         return (sock_net(sk)->ipv4.sysctl_tcp_fastopen & flag) ||
 317                tcp_sk(sk)->fastopen_no_cookie ||
 318                (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
 319 }
 320
 321 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
 322  * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
 323  * cookie request (foc->len == 0).
 324  */
 325 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
 326                               struct request_sock *req,
 327                               struct tcp_fastopen_cookie *foc,
 328                               const struct dst_entry *dst)
 329 {
 330         bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
 331         int tcp_fastopen = sock_net(sk)->ipv4.sysctl_tcp_fastopen;
 332         struct tcp_fastopen_cookie valid_foc = { .len = -1 };
 333         struct sock *child;
 334
 335         if (foc->len == 0) /* Client requests a cookie */
 336                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
 337
 338         if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
 339               (syn_data || foc->len >= 0) &&
 340               tcp_fastopen_queue_check(sk))) {
 341                 foc->len = -1;
 342                 return NULL;
 343         }
 344
 345         if (syn_data &&
 346             tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
 347                 goto fastopen;
 348
 349         if (foc->len >= 0 &&  /* Client presents or requests a cookie */
 350             tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc) &&
 351             foc->len == TCP_FASTOPEN_COOKIE_SIZE &&
 352             foc->len == valid_foc.len &&
 353             !memcmp(foc->val, valid_foc.val, foc->len)) {
 354                 /* Cookie is valid. Create a (full) child socket to accept
 355                  * the data in SYN before returning a SYN-ACK to ack the
 356                  * data. If we fail to create the socket, fall back and
 357                  * ack the ISN only but includes the same cookie.
 358                  *
 359                  * Note: Data-less SYN with valid cookie is allowed to send
 360                  * data in SYN_RECV state.
 361                  */
 362 fastopen:
 363                 child = tcp_fastopen_create_child(sk, skb, req);
 364                 if (child) {
 365                         foc->len = -1;
 366                         NET_INC_STATS(sock_net(sk),
 367                                       LINUX_MIB_TCPFASTOPENPASSIVE);
 368                         return child;
 369                 }
 370                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
 371         } else if (foc->len > 0) /* Client presents an invalid cookie */
 372                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
 373
 374         valid_foc.exp = foc->exp;
 375         *foc = valid_foc;
 376         return NULL;
 377 }
 378
 379 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
 380                                struct tcp_fastopen_cookie *cookie)
 381 {
 382         const struct dst_entry *dst;
 383
 384         tcp_fastopen_cache_get(sk, mss, cookie);
 385
 386         /* Firewall blackhole issue check */
 387         if (tcp_fastopen_active_should_disable(sk)) {
 388                 cookie->len = -1;
 389                 return false;
 390         }
 391
 392         dst = __sk_dst_get(sk);
 393
 394         if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
 395                 cookie->len = -1;
 396                 return true;
 397         }
 398         return cookie->len > 0;
 399 }
 400
 401 /* This function checks if we want to defer sending SYN until the first
 402  * write().  We defer under the following conditions:
 403  * 1. fastopen_connect sockopt is set
 404  * 2. we have a valid cookie
 405  * Return value: return true if we want to defer until application writes data
 406  *               return false if we want to send out SYN immediately
 407  */
 408 bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
 409 {
 410         struct tcp_fastopen_cookie cookie = { .len = 0 };
 411         struct tcp_sock *tp = tcp_sk(sk);
 412         u16 mss;
 413
 414         if (tp->fastopen_connect && !tp->fastopen_req) {
 415                 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
 416                         inet_sk(sk)->defer_connect = 1;
 417                         return true;
 418                 }
 419
 420                 /* Alloc fastopen_req in order for FO option to be included
 421                  * in SYN
 422                  */
 423                 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
 424                                            sk->sk_allocation);
 425                 if (tp->fastopen_req)
 426                         tp->fastopen_req->cookie = cookie;
 427                 else
 428                         *err = -ENOBUFS;
 429         }
 430         return false;
 431 }
 432 EXPORT_SYMBOL(tcp_fastopen_defer_connect);
 433
 434 /*
 435  * The following code block is to deal with middle box issues with TFO:
 436  * Middlebox firewall issues can potentially cause server's data being
 437  * blackholed after a successful 3WHS using TFO.
 438  * The proposed solution is to disable active TFO globally under the
 439  * following circumstances:
 440  *   1. client side TFO socket receives out of order FIN
 441  *   2. client side TFO socket receives out of order RST
 442  *   3. client side TFO socket has timed out three times consecutively during
 443  *      or after handshake
 444  * We disable active side TFO globally for 1hr at first. Then if it
 445  * happens again, we disable it for 2h, then 4h, 8h, ...
 446  * And we reset the timeout back to 1hr when we see a successful active
 447  * TFO connection with data exchanges.
 448  */
 449
 450 /* Disable active TFO and record current jiffies and
 451  * tfo_active_disable_times
 452  */
 453 void tcp_fastopen_active_disable(struct sock *sk)
 454 {
 455         struct net *net = sock_net(sk);
 456
 457         atomic_inc(&net->ipv4.tfo_active_disable_times);
 458         net->ipv4.tfo_active_disable_stamp = jiffies;
 459         NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
 460 }
 461
 462 /* Calculate timeout for tfo active disable
 463  * Return true if we are still in the active TFO disable period
 464  * Return false if timeout already expired and we should use active TFO
 465  */
 466 bool tcp_fastopen_active_should_disable(struct sock *sk)
 467 {
 468         unsigned int tfo_bh_timeout = sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout;
 469         int tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
 470         unsigned long timeout;
 471         int multiplier;
 472
 473         if (!tfo_da_times)
 474                 return false;
 475
 476         /* Limit timout to max: 2^6 * initial timeout */
 477         multiplier = 1 << min(tfo_da_times - 1, 6);
 478         timeout = multiplier * tfo_bh_timeout * HZ;
 479         if (time_before(jiffies, sock_net(sk)->ipv4.tfo_active_disable_stamp + timeout))
 480                 return true;
 481
 482         /* Mark check bit so we can check for successful active TFO
 483          * condition and reset tfo_active_disable_times
 484          */
 485         tcp_sk(sk)->syn_fastopen_ch = 1;
 486         return false;
 487 }
 488
 489 /* Disable active TFO if FIN is the only packet in the ofo queue
 490  * and no data is received.
 491  * Also check if we can reset tfo_active_disable_times if data is
 492  * received successfully on a marked active TFO sockets opened on
 493  * a non-loopback interface
 494  */
 495 void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
 496 {
 497         struct tcp_sock *tp = tcp_sk(sk);
 498         struct dst_entry *dst;
 499         struct sk_buff *skb;
 500
 501         if (!tp->syn_fastopen)
 502                 return;
 503
 504         if (!tp->data_segs_in) {
 505                 skb = skb_rb_first(&tp->out_of_order_queue);
 506                 if (skb && !skb_rb_next(skb)) {
 507                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
 508                                 tcp_fastopen_active_disable(sk);
 509                                 return;
 510                         }
 511                 }
 512         } else if (tp->syn_fastopen_ch &&
 513                    atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
 514                 dst = sk_dst_get(sk);
 515                 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
 516                         atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
 517                 dst_release(dst);
 518         }
 519 }
 520
 521 void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
 522 {
 523         u32 timeouts = inet_csk(sk)->icsk_retransmits;
 524         struct tcp_sock *tp = tcp_sk(sk);
 525
 526         /* Broken middle-boxes may black-hole Fast Open connection during or
 527          * even after the handshake. Be extremely conservative and pause
 528          * Fast Open globally after hitting the third consecutive timeout or
 529          * exceeding the configured timeout limit.
 530          */
 531         if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
 532             (timeouts == 2 || (timeouts < 2 && expired))) {
 533                 tcp_fastopen_active_disable(sk);
 534                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
 535         }
 536 }