1 #include <linux/err.h> 2 #include <linux/init.h> 3 #include <linux/kernel.h> 4 #include <linux/list.h> 5 #include <linux/tcp.h> 6 #include <linux/rcupdate.h> 7 #include <linux/rculist.h> 8 #include <net/inetpeer.h> 9 #include <net/tcp.h> 10 11 int sysctl_tcp_fastopen __read_mostly = TFO_CLIENT_ENABLE; 12 13 struct tcp_fastopen_context __rcu *tcp_fastopen_ctx; 14 15 static DEFINE_SPINLOCK(tcp_fastopen_ctx_lock); 16 17 void tcp_fastopen_init_key_once(bool publish) 18 { 19 static u8 key[TCP_FASTOPEN_KEY_LENGTH]; 20 21 /* tcp_fastopen_reset_cipher publishes the new context 22 * atomically, so we allow this race happening here. 23 * 24 * All call sites of tcp_fastopen_cookie_gen also check 25 * for a valid cookie, so this is an acceptable risk. 26 */ 27 if (net_get_random_once(key, sizeof(key)) && publish) 28 tcp_fastopen_reset_cipher(key, sizeof(key)); 29 } 30 31 static void tcp_fastopen_ctx_free(struct rcu_head *head) 32 { 33 struct tcp_fastopen_context *ctx = 34 container_of(head, struct tcp_fastopen_context, rcu); 35 crypto_free_cipher(ctx->tfm); 36 kfree(ctx); 37 } 38 39 int tcp_fastopen_reset_cipher(void *key, unsigned int len) 40 { 41 int err; 42 struct tcp_fastopen_context *ctx, *octx; 43 44 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); 45 if (!ctx) 46 return -ENOMEM; 47 ctx->tfm = crypto_alloc_cipher("aes", 0, 0); 48 49 if (IS_ERR(ctx->tfm)) { 50 err = PTR_ERR(ctx->tfm); 51 error: kfree(ctx); 52 pr_err("TCP: TFO aes cipher alloc error: %d\n", err); 53 return err; 54 } 55 err = crypto_cipher_setkey(ctx->tfm, key, len); 56 if (err) { 57 pr_err("TCP: TFO cipher key error: %d\n", err); 58 crypto_free_cipher(ctx->tfm); 59 goto error; 60 } 61 memcpy(ctx->key, key, len); 62 63 spin_lock(&tcp_fastopen_ctx_lock); 64 65 octx = rcu_dereference_protected(tcp_fastopen_ctx, 66 lockdep_is_held(&tcp_fastopen_ctx_lock)); 67 rcu_assign_pointer(tcp_fastopen_ctx, ctx); 68 spin_unlock(&tcp_fastopen_ctx_lock); 69 70 if (octx) 71 call_rcu(&octx->rcu, tcp_fastopen_ctx_free); 72 return err; 73 } 74 75 static bool __tcp_fastopen_cookie_gen(const void *path, 76 struct tcp_fastopen_cookie *foc) 77 { 78 struct tcp_fastopen_context *ctx; 79 bool ok = false; 80 81 rcu_read_lock(); 82 ctx = rcu_dereference(tcp_fastopen_ctx); 83 if (ctx) { 84 crypto_cipher_encrypt_one(ctx->tfm, foc->val, path); 85 foc->len = TCP_FASTOPEN_COOKIE_SIZE; 86 ok = true; 87 } 88 rcu_read_unlock(); 89 return ok; 90 } 91 92 /* Generate the fastopen cookie by doing aes128 encryption on both 93 * the source and destination addresses. Pad 0s for IPv4 or IPv4-mapped-IPv6 94 * addresses. For the longer IPv6 addresses use CBC-MAC. 95 * 96 * XXX (TFO) - refactor when TCP_FASTOPEN_COOKIE_SIZE != AES_BLOCK_SIZE. 97 */ 98 static bool tcp_fastopen_cookie_gen(struct request_sock *req, 99 struct sk_buff *syn, 100 struct tcp_fastopen_cookie *foc) 101 { 102 if (req->rsk_ops->family == AF_INET) { 103 const struct iphdr *iph = ip_hdr(syn); 104 105 __be32 path[4] = { iph->saddr, iph->daddr, 0, 0 }; 106 return __tcp_fastopen_cookie_gen(path, foc); 107 } 108 109 #if IS_ENABLED(CONFIG_IPV6) 110 if (req->rsk_ops->family == AF_INET6) { 111 const struct ipv6hdr *ip6h = ipv6_hdr(syn); 112 struct tcp_fastopen_cookie tmp; 113 114 if (__tcp_fastopen_cookie_gen(&ip6h->saddr, &tmp)) { 115 struct in6_addr *buf = (struct in6_addr *) tmp.val; 116 int i; 117 118 for (i = 0; i < 4; i++) 119 buf->s6_addr32[i] ^= ip6h->daddr.s6_addr32[i]; 120 return __tcp_fastopen_cookie_gen(buf, foc); 121 } 122 } 123 #endif 124 return false; 125 } 126 127 static bool tcp_fastopen_create_child(struct sock *sk, 128 struct sk_buff *skb, 129 struct dst_entry *dst, 130 struct request_sock *req) 131 { 132 struct tcp_sock *tp; 133 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 134 struct sock *child; 135 u32 end_seq; 136 137 req->num_retrans = 0; 138 req->num_timeout = 0; 139 req->sk = NULL; 140 141 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL); 142 if (!child) 143 return false; 144 145 spin_lock(&queue->fastopenq->lock); 146 queue->fastopenq->qlen++; 147 spin_unlock(&queue->fastopenq->lock); 148 149 /* Initialize the child socket. Have to fix some values to take 150 * into account the child is a Fast Open socket and is created 151 * only out of the bits carried in the SYN packet. 152 */ 153 tp = tcp_sk(child); 154 155 tp->fastopen_rsk = req; 156 tcp_rsk(req)->tfo_listener = true; 157 158 /* RFC1323: The window in SYN & SYN/ACK segments is never 159 * scaled. So correct it appropriately. 160 */ 161 tp->snd_wnd = ntohs(tcp_hdr(skb)->window); 162 163 /* Activate the retrans timer so that SYNACK can be retransmitted. 164 * The request socket is not added to the SYN table of the parent 165 * because it's been added to the accept queue directly. 166 */ 167 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS, 168 TCP_TIMEOUT_INIT, TCP_RTO_MAX); 169 170 atomic_set(&req->rsk_refcnt, 1); 171 /* Add the child socket directly into the accept queue */ 172 inet_csk_reqsk_queue_add(sk, req, child); 173 174 /* Now finish processing the fastopen child socket. */ 175 inet_csk(child)->icsk_af_ops->rebuild_header(child); 176 tcp_init_congestion_control(child); 177 tcp_mtup_init(child); 178 tcp_init_metrics(child); 179 tcp_init_buffer_space(child); 180 181 /* Queue the data carried in the SYN packet. We need to first 182 * bump skb's refcnt because the caller will attempt to free it. 183 * Note that IPv6 might also have used skb_get() trick 184 * in tcp_v6_conn_request() to keep this SYN around (treq->pktopts) 185 * So we need to eventually get a clone of the packet, 186 * before inserting it in sk_receive_queue. 187 * 188 * XXX (TFO) - we honor a zero-payload TFO request for now, 189 * (any reason not to?) but no need to queue the skb since 190 * there is no data. How about SYN+FIN? 191 */ 192 end_seq = TCP_SKB_CB(skb)->end_seq; 193 if (end_seq != TCP_SKB_CB(skb)->seq + 1) { 194 struct sk_buff *skb2; 195 196 if (unlikely(skb_shared(skb))) 197 skb2 = skb_clone(skb, GFP_ATOMIC); 198 else 199 skb2 = skb_get(skb); 200 201 if (likely(skb2)) { 202 skb_dst_drop(skb2); 203 __skb_pull(skb2, tcp_hdrlen(skb)); 204 skb_set_owner_r(skb2, child); 205 __skb_queue_tail(&child->sk_receive_queue, skb2); 206 tp->syn_data_acked = 1; 207 208 /* u64_stats_update_begin(&tp->syncp) not needed here, 209 * as we certainly are not changing upper 32bit value (0) 210 */ 211 tp->bytes_received = end_seq - TCP_SKB_CB(skb)->seq - 1; 212 } else { 213 end_seq = TCP_SKB_CB(skb)->seq + 1; 214 } 215 } 216 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt = end_seq; 217 sk->sk_data_ready(sk); 218 bh_unlock_sock(child); 219 sock_put(child); 220 WARN_ON(!req->sk); 221 return true; 222 } 223 224 static bool tcp_fastopen_queue_check(struct sock *sk) 225 { 226 struct fastopen_queue *fastopenq; 227 228 /* Make sure the listener has enabled fastopen, and we don't 229 * exceed the max # of pending TFO requests allowed before trying 230 * to validating the cookie in order to avoid burning CPU cycles 231 * unnecessarily. 232 * 233 * XXX (TFO) - The implication of checking the max_qlen before 234 * processing a cookie request is that clients can't differentiate 235 * between qlen overflow causing Fast Open to be disabled 236 * temporarily vs a server not supporting Fast Open at all. 237 */ 238 fastopenq = inet_csk(sk)->icsk_accept_queue.fastopenq; 239 if (!fastopenq || fastopenq->max_qlen == 0) 240 return false; 241 242 if (fastopenq->qlen >= fastopenq->max_qlen) { 243 struct request_sock *req1; 244 spin_lock(&fastopenq->lock); 245 req1 = fastopenq->rskq_rst_head; 246 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) { 247 spin_unlock(&fastopenq->lock); 248 NET_INC_STATS_BH(sock_net(sk), 249 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW); 250 return false; 251 } 252 fastopenq->rskq_rst_head = req1->dl_next; 253 fastopenq->qlen--; 254 spin_unlock(&fastopenq->lock); 255 reqsk_put(req1); 256 } 257 return true; 258 } 259 260 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc) 261 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open 262 * cookie request (foc->len == 0). 263 */ 264 bool tcp_try_fastopen(struct sock *sk, struct sk_buff *skb, 265 struct request_sock *req, 266 struct tcp_fastopen_cookie *foc, 267 struct dst_entry *dst) 268 { 269 struct tcp_fastopen_cookie valid_foc = { .len = -1 }; 270 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1; 271 272 if (foc->len == 0) /* Client requests a cookie */ 273 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD); 274 275 if (!((sysctl_tcp_fastopen & TFO_SERVER_ENABLE) && 276 (syn_data || foc->len >= 0) && 277 tcp_fastopen_queue_check(sk))) { 278 foc->len = -1; 279 return false; 280 } 281 282 if (syn_data && (sysctl_tcp_fastopen & TFO_SERVER_COOKIE_NOT_REQD)) 283 goto fastopen; 284 285 if (foc->len >= 0 && /* Client presents or requests a cookie */ 286 tcp_fastopen_cookie_gen(req, skb, &valid_foc) && 287 foc->len == TCP_FASTOPEN_COOKIE_SIZE && 288 foc->len == valid_foc.len && 289 !memcmp(foc->val, valid_foc.val, foc->len)) { 290 /* Cookie is valid. Create a (full) child socket to accept 291 * the data in SYN before returning a SYN-ACK to ack the 292 * data. If we fail to create the socket, fall back and 293 * ack the ISN only but includes the same cookie. 294 * 295 * Note: Data-less SYN with valid cookie is allowed to send 296 * data in SYN_RECV state. 297 */ 298 fastopen: 299 if (tcp_fastopen_create_child(sk, skb, dst, req)) { 300 foc->len = -1; 301 NET_INC_STATS_BH(sock_net(sk), 302 LINUX_MIB_TCPFASTOPENPASSIVE); 303 return true; 304 } 305 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL); 306 } else if (foc->len > 0) /* Client presents an invalid cookie */ 307 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL); 308 309 valid_foc.exp = foc->exp; 310 *foc = valid_foc; 311 return false; 312 } 313 EXPORT_SYMBOL(tcp_try_fastopen); 314