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 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 225 NULL, &own_req); 226 if (!child) 227 return NULL; 228 229 spin_lock(&queue->fastopenq.lock); 230 queue->fastopenq.qlen++; 231 spin_unlock(&queue->fastopenq.lock); 232 233 /* Initialize the child socket. Have to fix some values to take 234 * into account the child is a Fast Open socket and is created 235 * only out of the bits carried in the SYN packet. 236 */ 237 tp = tcp_sk(child); 238 239 tp->fastopen_rsk = req; 240 tcp_rsk(req)->tfo_listener = true; 241 242 /* RFC1323: The window in SYN & SYN/ACK segments is never 243 * scaled. So correct it appropriately. 244 */ 245 tp->snd_wnd = ntohs(tcp_hdr(skb)->window); 246 tp->max_window = tp->snd_wnd; 247 248 /* Activate the retrans timer so that SYNACK can be retransmitted. 249 * The request socket is not added to the ehash 250 * because it's been added to the accept queue directly. 251 */ 252 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS, 253 TCP_TIMEOUT_INIT, TCP_RTO_MAX); 254 255 refcount_set(&req->rsk_refcnt, 2); 256 257 /* Now finish processing the fastopen child socket. */ 258 tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB); 259 260 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 261 262 tcp_fastopen_add_skb(child, skb); 263 264 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt; 265 tp->rcv_wup = tp->rcv_nxt; 266 /* tcp_conn_request() is sending the SYNACK, 267 * and queues the child into listener accept queue. 268 */ 269 return child; 270 } 271 272 static bool tcp_fastopen_queue_check(struct sock *sk) 273 { 274 struct fastopen_queue *fastopenq; 275 276 /* Make sure the listener has enabled fastopen, and we don't 277 * exceed the max # of pending TFO requests allowed before trying 278 * to validating the cookie in order to avoid burning CPU cycles 279 * unnecessarily. 280 * 281 * XXX (TFO) - The implication of checking the max_qlen before 282 * processing a cookie request is that clients can't differentiate 283 * between qlen overflow causing Fast Open to be disabled 284 * temporarily vs a server not supporting Fast Open at all. 285 */ 286 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq; 287 if (fastopenq->max_qlen == 0) 288 return false; 289 290 if (fastopenq->qlen >= fastopenq->max_qlen) { 291 struct request_sock *req1; 292 spin_lock(&fastopenq->lock); 293 req1 = fastopenq->rskq_rst_head; 294 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) { 295 __NET_INC_STATS(sock_net(sk), 296 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW); 297 spin_unlock(&fastopenq->lock); 298 return false; 299 } 300 fastopenq->rskq_rst_head = req1->dl_next; 301 fastopenq->qlen--; 302 spin_unlock(&fastopenq->lock); 303 reqsk_put(req1); 304 } 305 return true; 306 } 307 308 static bool tcp_fastopen_no_cookie(const struct sock *sk, 309 const struct dst_entry *dst, 310 int flag) 311 { 312 return (sock_net(sk)->ipv4.sysctl_tcp_fastopen & flag) || 313 tcp_sk(sk)->fastopen_no_cookie || 314 (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE)); 315 } 316 317 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc) 318 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open 319 * cookie request (foc->len == 0). 320 */ 321 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb, 322 struct request_sock *req, 323 struct tcp_fastopen_cookie *foc, 324 const struct dst_entry *dst) 325 { 326 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1; 327 int tcp_fastopen = sock_net(sk)->ipv4.sysctl_tcp_fastopen; 328 struct tcp_fastopen_cookie valid_foc = { .len = -1 }; 329 struct sock *child; 330 331 if (foc->len == 0) /* Client requests a cookie */ 332 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD); 333 334 if (!((tcp_fastopen & TFO_SERVER_ENABLE) && 335 (syn_data || foc->len >= 0) && 336 tcp_fastopen_queue_check(sk))) { 337 foc->len = -1; 338 return NULL; 339 } 340 341 if (syn_data && 342 tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD)) 343 goto fastopen; 344 345 if (foc->len >= 0 && /* Client presents or requests a cookie */ 346 tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc) && 347 foc->len == TCP_FASTOPEN_COOKIE_SIZE && 348 foc->len == valid_foc.len && 349 !memcmp(foc->val, valid_foc.val, foc->len)) { 350 /* Cookie is valid. Create a (full) child socket to accept 351 * the data in SYN before returning a SYN-ACK to ack the 352 * data. If we fail to create the socket, fall back and 353 * ack the ISN only but includes the same cookie. 354 * 355 * Note: Data-less SYN with valid cookie is allowed to send 356 * data in SYN_RECV state. 357 */ 358 fastopen: 359 child = tcp_fastopen_create_child(sk, skb, req); 360 if (child) { 361 foc->len = -1; 362 NET_INC_STATS(sock_net(sk), 363 LINUX_MIB_TCPFASTOPENPASSIVE); 364 return child; 365 } 366 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL); 367 } else if (foc->len > 0) /* Client presents an invalid cookie */ 368 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL); 369 370 valid_foc.exp = foc->exp; 371 *foc = valid_foc; 372 return NULL; 373 } 374 375 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss, 376 struct tcp_fastopen_cookie *cookie) 377 { 378 const struct dst_entry *dst; 379 380 tcp_fastopen_cache_get(sk, mss, cookie); 381 382 /* Firewall blackhole issue check */ 383 if (tcp_fastopen_active_should_disable(sk)) { 384 cookie->len = -1; 385 return false; 386 } 387 388 dst = __sk_dst_get(sk); 389 390 if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) { 391 cookie->len = -1; 392 return true; 393 } 394 return cookie->len > 0; 395 } 396 397 /* This function checks if we want to defer sending SYN until the first 398 * write(). We defer under the following conditions: 399 * 1. fastopen_connect sockopt is set 400 * 2. we have a valid cookie 401 * Return value: return true if we want to defer until application writes data 402 * return false if we want to send out SYN immediately 403 */ 404 bool tcp_fastopen_defer_connect(struct sock *sk, int *err) 405 { 406 struct tcp_fastopen_cookie cookie = { .len = 0 }; 407 struct tcp_sock *tp = tcp_sk(sk); 408 u16 mss; 409 410 if (tp->fastopen_connect && !tp->fastopen_req) { 411 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) { 412 inet_sk(sk)->defer_connect = 1; 413 return true; 414 } 415 416 /* Alloc fastopen_req in order for FO option to be included 417 * in SYN 418 */ 419 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req), 420 sk->sk_allocation); 421 if (tp->fastopen_req) 422 tp->fastopen_req->cookie = cookie; 423 else 424 *err = -ENOBUFS; 425 } 426 return false; 427 } 428 EXPORT_SYMBOL(tcp_fastopen_defer_connect); 429 430 /* 431 * The following code block is to deal with middle box issues with TFO: 432 * Middlebox firewall issues can potentially cause server's data being 433 * blackholed after a successful 3WHS using TFO. 434 * The proposed solution is to disable active TFO globally under the 435 * following circumstances: 436 * 1. client side TFO socket receives out of order FIN 437 * 2. client side TFO socket receives out of order RST 438 * 3. client side TFO socket has timed out three times consecutively during 439 * or after handshake 440 * We disable active side TFO globally for 1hr at first. Then if it 441 * happens again, we disable it for 2h, then 4h, 8h, ... 442 * And we reset the timeout back to 1hr when we see a successful active 443 * TFO connection with data exchanges. 444 */ 445 446 /* Disable active TFO and record current jiffies and 447 * tfo_active_disable_times 448 */ 449 void tcp_fastopen_active_disable(struct sock *sk) 450 { 451 struct net *net = sock_net(sk); 452 453 atomic_inc(&net->ipv4.tfo_active_disable_times); 454 net->ipv4.tfo_active_disable_stamp = jiffies; 455 NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE); 456 } 457 458 /* Calculate timeout for tfo active disable 459 * Return true if we are still in the active TFO disable period 460 * Return false if timeout already expired and we should use active TFO 461 */ 462 bool tcp_fastopen_active_should_disable(struct sock *sk) 463 { 464 unsigned int tfo_bh_timeout = sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout; 465 int tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times); 466 unsigned long timeout; 467 int multiplier; 468 469 if (!tfo_da_times) 470 return false; 471 472 /* Limit timout to max: 2^6 * initial timeout */ 473 multiplier = 1 << min(tfo_da_times - 1, 6); 474 timeout = multiplier * tfo_bh_timeout * HZ; 475 if (time_before(jiffies, sock_net(sk)->ipv4.tfo_active_disable_stamp + timeout)) 476 return true; 477 478 /* Mark check bit so we can check for successful active TFO 479 * condition and reset tfo_active_disable_times 480 */ 481 tcp_sk(sk)->syn_fastopen_ch = 1; 482 return false; 483 } 484 485 /* Disable active TFO if FIN is the only packet in the ofo queue 486 * and no data is received. 487 * Also check if we can reset tfo_active_disable_times if data is 488 * received successfully on a marked active TFO sockets opened on 489 * a non-loopback interface 490 */ 491 void tcp_fastopen_active_disable_ofo_check(struct sock *sk) 492 { 493 struct tcp_sock *tp = tcp_sk(sk); 494 struct dst_entry *dst; 495 struct sk_buff *skb; 496 497 if (!tp->syn_fastopen) 498 return; 499 500 if (!tp->data_segs_in) { 501 skb = skb_rb_first(&tp->out_of_order_queue); 502 if (skb && !skb_rb_next(skb)) { 503 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { 504 tcp_fastopen_active_disable(sk); 505 return; 506 } 507 } 508 } else if (tp->syn_fastopen_ch && 509 atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) { 510 dst = sk_dst_get(sk); 511 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK))) 512 atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0); 513 dst_release(dst); 514 } 515 } 516 517 void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired) 518 { 519 u32 timeouts = inet_csk(sk)->icsk_retransmits; 520 struct tcp_sock *tp = tcp_sk(sk); 521 522 /* Broken middle-boxes may black-hole Fast Open connection during or 523 * even after the handshake. Be extremely conservative and pause 524 * Fast Open globally after hitting the third consecutive timeout or 525 * exceeding the configured timeout limit. 526 */ 527 if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) && 528 (timeouts == 2 || (timeouts < 2 && expired))) { 529 tcp_fastopen_active_disable(sk); 530 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL); 531 } 532 } 533