net/ipv4/tcp_fastopen.c

   1 #include <linux/crypto.h>
   2 #include <linux/err.h>
   3 #include <linux/init.h>
   4 #include <linux/kernel.h>
   5 #include <linux/list.h>
   6 #include <linux/tcp.h>
   7 #include <linux/rcupdate.h>
   8 #include <linux/rculist.h>
   9 #include <net/inetpeer.h>
  10 #include <net/tcp.h>
  11
  12 int sysctl_tcp_fastopen __read_mostly = TFO_CLIENT_ENABLE;
  13
  14 struct tcp_fastopen_context __rcu *tcp_fastopen_ctx;
  15
  16 static DEFINE_SPINLOCK(tcp_fastopen_ctx_lock);
  17
  18 void tcp_fastopen_init_key_once(bool publish)
  19 {
  20         static u8 key[TCP_FASTOPEN_KEY_LENGTH];
  21
  22         /* tcp_fastopen_reset_cipher publishes the new context
  23          * atomically, so we allow this race happening here.
  24          *
  25          * All call sites of tcp_fastopen_cookie_gen also check
  26          * for a valid cookie, so this is an acceptable risk.
  27          */
  28         if (net_get_random_once(key, sizeof(key)) && publish)
  29                 tcp_fastopen_reset_cipher(key, sizeof(key));
  30 }
  31
  32 static void tcp_fastopen_ctx_free(struct rcu_head *head)
  33 {
  34         struct tcp_fastopen_context *ctx =
  35             container_of(head, struct tcp_fastopen_context, rcu);
  36         crypto_free_cipher(ctx->tfm);
  37         kfree(ctx);
  38 }
  39
  40 int tcp_fastopen_reset_cipher(void *key, unsigned int len)
  41 {
  42         int err;
  43         struct tcp_fastopen_context *ctx, *octx;
  44
  45         ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
  46         if (!ctx)
  47                 return -ENOMEM;
  48         ctx->tfm = crypto_alloc_cipher("aes", 0, 0);
  49
  50         if (IS_ERR(ctx->tfm)) {
  51                 err = PTR_ERR(ctx->tfm);
  52 error:          kfree(ctx);
  53                 pr_err("TCP: TFO aes cipher alloc error: %d\n", err);
  54                 return err;
  55         }
  56         err = crypto_cipher_setkey(ctx->tfm, key, len);
  57         if (err) {
  58                 pr_err("TCP: TFO cipher key error: %d\n", err);
  59                 crypto_free_cipher(ctx->tfm);
  60                 goto error;
  61         }
  62         memcpy(ctx->key, key, len);
  63
  64         spin_lock(&tcp_fastopen_ctx_lock);
  65
  66         octx = rcu_dereference_protected(tcp_fastopen_ctx,
  67                                 lockdep_is_held(&tcp_fastopen_ctx_lock));
  68         rcu_assign_pointer(tcp_fastopen_ctx, ctx);
  69         spin_unlock(&tcp_fastopen_ctx_lock);
  70
  71         if (octx)
  72                 call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
  73         return err;
  74 }
  75
  76 static bool __tcp_fastopen_cookie_gen(const void *path,
  77                                       struct tcp_fastopen_cookie *foc)
  78 {
  79         struct tcp_fastopen_context *ctx;
  80         bool ok = false;
  81
  82         rcu_read_lock();
  83         ctx = rcu_dereference(tcp_fastopen_ctx);
  84         if (ctx) {
  85                 crypto_cipher_encrypt_one(ctx->tfm, foc->val, path);
  86                 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
  87                 ok = true;
  88         }
  89         rcu_read_unlock();
  90         return ok;
  91 }
  92
  93 /* Generate the fastopen cookie by doing aes128 encryption on both
  94  * the source and destination addresses. Pad 0s for IPv4 or IPv4-mapped-IPv6
  95  * addresses. For the longer IPv6 addresses use CBC-MAC.
  96  *
  97  * XXX (TFO) - refactor when TCP_FASTOPEN_COOKIE_SIZE != AES_BLOCK_SIZE.
  98  */
  99 static bool tcp_fastopen_cookie_gen(struct request_sock *req,
 100                                     struct sk_buff *syn,
 101                                     struct tcp_fastopen_cookie *foc)
 102 {
 103         if (req->rsk_ops->family == AF_INET) {
 104                 const struct iphdr *iph = ip_hdr(syn);
 105
 106                 __be32 path[4] = { iph->saddr, iph->daddr, 0, 0 };
 107                 return __tcp_fastopen_cookie_gen(path, foc);
 108         }
 109
 110 #if IS_ENABLED(CONFIG_IPV6)
 111         if (req->rsk_ops->family == AF_INET6) {
 112                 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
 113                 struct tcp_fastopen_cookie tmp;
 114
 115                 if (__tcp_fastopen_cookie_gen(&ip6h->saddr, &tmp)) {
 116                         struct in6_addr *buf = (struct in6_addr *) tmp.val;
 117                         int i;
 118
 119                         for (i = 0; i < 4; i++)
 120                                 buf->s6_addr32[i] ^= ip6h->daddr.s6_addr32[i];
 121                         return __tcp_fastopen_cookie_gen(buf, foc);
 122                 }
 123         }
 124 #endif
 125         return false;
 126 }
 127
 128
 129 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
 130  * queue this additional data / FIN.
 131  */
 132 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
 133 {
 134         struct tcp_sock *tp = tcp_sk(sk);
 135
 136         if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
 137                 return;
 138
 139         skb = skb_clone(skb, GFP_ATOMIC);
 140         if (!skb)
 141                 return;
 142
 143         skb_dst_drop(skb);
 144         /* segs_in has been initialized to 1 in tcp_create_openreq_child().
 145          * Hence, reset segs_in to 0 before calling tcp_segs_in()
 146          * to avoid double counting.  Also, tcp_segs_in() expects
 147          * skb->len to include the tcp_hdrlen.  Hence, it should
 148          * be called before __skb_pull().
 149          */
 150         tp->segs_in = 0;
 151         tcp_segs_in(tp, skb);
 152         __skb_pull(skb, tcp_hdrlen(skb));
 153         sk_forced_mem_schedule(sk, skb->truesize);
 154         skb_set_owner_r(skb, sk);
 155
 156         TCP_SKB_CB(skb)->seq++;
 157         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
 158
 159         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
 160         __skb_queue_tail(&sk->sk_receive_queue, skb);
 161         tp->syn_data_acked = 1;
 162
 163         /* u64_stats_update_begin(&tp->syncp) not needed here,
 164          * as we certainly are not changing upper 32bit value (0)
 165          */
 166         tp->bytes_received = skb->len;
 167
 168         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
 169                 tcp_fin(sk);
 170 }
 171
 172 static struct sock *tcp_fastopen_create_child(struct sock *sk,
 173                                               struct sk_buff *skb,
 174                                               struct dst_entry *dst,
 175                                               struct request_sock *req)
 176 {
 177         struct tcp_sock *tp;
 178         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 179         struct sock *child;
 180         bool own_req;
 181
 182         req->num_retrans = 0;
 183         req->num_timeout = 0;
 184         req->sk = NULL;
 185
 186         child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
 187                                                          NULL, &own_req);
 188         if (!child)
 189                 return NULL;
 190
 191         spin_lock(&queue->fastopenq.lock);
 192         queue->fastopenq.qlen++;
 193         spin_unlock(&queue->fastopenq.lock);
 194
 195         /* Initialize the child socket. Have to fix some values to take
 196          * into account the child is a Fast Open socket and is created
 197          * only out of the bits carried in the SYN packet.
 198          */
 199         tp = tcp_sk(child);
 200
 201         tp->fastopen_rsk = req;
 202         tcp_rsk(req)->tfo_listener = true;
 203
 204         /* RFC1323: The window in SYN & SYN/ACK segments is never
 205          * scaled. So correct it appropriately.
 206          */
 207         tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
 208
 209         /* Activate the retrans timer so that SYNACK can be retransmitted.
 210          * The request socket is not added to the ehash
 211          * because it's been added to the accept queue directly.
 212          */
 213         inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
 214                                   TCP_TIMEOUT_INIT, TCP_RTO_MAX);
 215
 216         atomic_set(&req->rsk_refcnt, 2);
 217
 218         /* Now finish processing the fastopen child socket. */
 219         inet_csk(child)->icsk_af_ops->rebuild_header(child);
 220         tcp_init_congestion_control(child);
 221         tcp_mtup_init(child);
 222         tcp_init_metrics(child);
 223         tcp_init_buffer_space(child);
 224
 225         tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
 226
 227         tcp_fastopen_add_skb(child, skb);
 228
 229         tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
 230         tp->rcv_wup = tp->rcv_nxt;
 231         /* tcp_conn_request() is sending the SYNACK,
 232          * and queues the child into listener accept queue.
 233          */
 234         return child;
 235 }
 236
 237 static bool tcp_fastopen_queue_check(struct sock *sk)
 238 {
 239         struct fastopen_queue *fastopenq;
 240
 241         /* Make sure the listener has enabled fastopen, and we don't
 242          * exceed the max # of pending TFO requests allowed before trying
 243          * to validating the cookie in order to avoid burning CPU cycles
 244          * unnecessarily.
 245          *
 246          * XXX (TFO) - The implication of checking the max_qlen before
 247          * processing a cookie request is that clients can't differentiate
 248          * between qlen overflow causing Fast Open to be disabled
 249          * temporarily vs a server not supporting Fast Open at all.
 250          */
 251         fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
 252         if (fastopenq->max_qlen == 0)
 253                 return false;
 254
 255         if (fastopenq->qlen >= fastopenq->max_qlen) {
 256                 struct request_sock *req1;
 257                 spin_lock(&fastopenq->lock);
 258                 req1 = fastopenq->rskq_rst_head;
 259                 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
 260                         __NET_INC_STATS(sock_net(sk),
 261                                         LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
 262                         spin_unlock(&fastopenq->lock);
 263                         return false;
 264                 }
 265                 fastopenq->rskq_rst_head = req1->dl_next;
 266                 fastopenq->qlen--;
 267                 spin_unlock(&fastopenq->lock);
 268                 reqsk_put(req1);
 269         }
 270         return true;
 271 }
 272
 273 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
 274  * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
 275  * cookie request (foc->len == 0).
 276  */
 277 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
 278                               struct request_sock *req,
 279                               struct tcp_fastopen_cookie *foc,
 280                               struct dst_entry *dst)
 281 {
 282         struct tcp_fastopen_cookie valid_foc = { .len = -1 };
 283         bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
 284         struct sock *child;
 285
 286         if (foc->len == 0) /* Client requests a cookie */
 287                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
 288
 289         if (!((sysctl_tcp_fastopen & TFO_SERVER_ENABLE) &&
 290               (syn_data || foc->len >= 0) &&
 291               tcp_fastopen_queue_check(sk))) {
 292                 foc->len = -1;
 293                 return NULL;
 294         }
 295
 296         if (syn_data && (sysctl_tcp_fastopen & TFO_SERVER_COOKIE_NOT_REQD))
 297                 goto fastopen;
 298
 299         if (foc->len >= 0 &&  /* Client presents or requests a cookie */
 300             tcp_fastopen_cookie_gen(req, skb, &valid_foc) &&
 301             foc->len == TCP_FASTOPEN_COOKIE_SIZE &&
 302             foc->len == valid_foc.len &&
 303             !memcmp(foc->val, valid_foc.val, foc->len)) {
 304                 /* Cookie is valid. Create a (full) child socket to accept
 305                  * the data in SYN before returning a SYN-ACK to ack the
 306                  * data. If we fail to create the socket, fall back and
 307                  * ack the ISN only but includes the same cookie.
 308                  *
 309                  * Note: Data-less SYN with valid cookie is allowed to send
 310                  * data in SYN_RECV state.
 311                  */
 312 fastopen:
 313                 child = tcp_fastopen_create_child(sk, skb, dst, req);
 314                 if (child) {
 315                         foc->len = -1;
 316                         NET_INC_STATS(sock_net(sk),
 317                                       LINUX_MIB_TCPFASTOPENPASSIVE);
 318                         return child;
 319                 }
 320                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
 321         } else if (foc->len > 0) /* Client presents an invalid cookie */
 322                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
 323
 324         valid_foc.exp = foc->exp;
 325         *foc = valid_foc;
 326         return NULL;
 327 }