net/dccp/output.c

   1 /*
   2  *  net/dccp/output.c
   3  *
   4  *  An implementation of the DCCP protocol
   5  *  Arnaldo Carvalho de Melo <acme@conectiva.com.br>
   6  *
   7  *      This program is free software; you can redistribute it and/or
   8  *      modify it under the terms of the GNU General Public License
   9  *      as published by the Free Software Foundation; either version
  10  *      2 of the License, or (at your option) any later version.
  11  */
  12
  13 #include <linux/config.h>
  14 #include <linux/dccp.h>
  15 #include <linux/skbuff.h>
  16
  17 #include <net/sock.h>
  18
  19 #include "ackvec.h"
  20 #include "ccid.h"
  21 #include "dccp.h"
  22
  23 static inline void dccp_event_ack_sent(struct sock *sk)
  24 {
  25         inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
  26 }
  27
  28 /*
  29  * All SKB's seen here are completely headerless. It is our
  30  * job to build the DCCP header, and pass the packet down to
  31  * IP so it can do the same plus pass the packet off to the
  32  * device.
  33  */
  34 int dccp_transmit_skb(struct sock *sk, struct sk_buff *skb)
  35 {
  36         if (likely(skb != NULL)) {
  37                 const struct inet_sock *inet = inet_sk(sk);
  38                 struct dccp_sock *dp = dccp_sk(sk);
  39                 struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
  40                 struct dccp_hdr *dh;
  41                 /* XXX For now we're using only 48 bits sequence numbers */
  42                 const int dccp_header_size = sizeof(*dh) +
  43                                              sizeof(struct dccp_hdr_ext) +
  44                                           dccp_packet_hdr_len(dcb->dccpd_type);
  45                 int err, set_ack = 1;
  46                 u64 ackno = dp->dccps_gsr;
  47
  48                 dccp_inc_seqno(&dp->dccps_gss);
  49
  50                 switch (dcb->dccpd_type) {
  51                 case DCCP_PKT_DATA:
  52                         set_ack = 0;
  53                         break;
  54                 case DCCP_PKT_SYNC:
  55                 case DCCP_PKT_SYNCACK:
  56                         ackno = dcb->dccpd_seq;
  57                         break;
  58                 }
  59
  60                 dcb->dccpd_seq = dp->dccps_gss;
  61                 dccp_insert_options(sk, skb);
  62
  63                 skb->h.raw = skb_push(skb, dccp_header_size);
  64                 dh = dccp_hdr(skb);
  65
  66                 if (!skb->sk)
  67                         skb_set_owner_w(skb, sk);
  68
  69                 /* Build DCCP header and checksum it. */
  70                 memset(dh, 0, dccp_header_size);
  71                 dh->dccph_type  = dcb->dccpd_type;
  72                 dh->dccph_sport = inet->sport;
  73                 dh->dccph_dport = inet->dport;
  74                 dh->dccph_doff  = (dccp_header_size + dcb->dccpd_opt_len) / 4;
  75                 dh->dccph_ccval = dcb->dccpd_ccval;
  76                 /* XXX For now we're using only 48 bits sequence numbers */
  77                 dh->dccph_x     = 1;
  78
  79                 dp->dccps_awh = dp->dccps_gss;
  80                 dccp_hdr_set_seq(dh, dp->dccps_gss);
  81                 if (set_ack)
  82                         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), ackno);
  83
  84                 switch (dcb->dccpd_type) {
  85                 case DCCP_PKT_REQUEST:
  86                         dccp_hdr_request(skb)->dccph_req_service =
  87                                                         dp->dccps_service;
  88                         break;
  89                 case DCCP_PKT_RESET:
  90                         dccp_hdr_reset(skb)->dccph_reset_code =
  91                                                         dcb->dccpd_reset_code;
  92                         break;
  93                 }
  94
  95                 dh->dccph_checksum = dccp_v4_checksum(skb, inet->saddr,
  96                                                       inet->daddr);
  97
  98                 if (set_ack)
  99                         dccp_event_ack_sent(sk);
 100
 101                 DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
 102
 103                 err = ip_queue_xmit(skb, 0);
 104                 if (err <= 0)
 105                         return err;
 106
 107                 /* NET_XMIT_CN is special. It does not guarantee,
 108                  * that this packet is lost. It tells that device
 109                  * is about to start to drop packets or already
 110                  * drops some packets of the same priority and
 111                  * invokes us to send less aggressively.
 112                  */
 113                 return err == NET_XMIT_CN ? 0 : err;
 114         }
 115         return -ENOBUFS;
 116 }
 117
 118 unsigned int dccp_sync_mss(struct sock *sk, u32 pmtu)
 119 {
 120         struct dccp_sock *dp = dccp_sk(sk);
 121         int mss_now;
 122
 123         /*
 124          * FIXME: we really should be using the af_specific thing to support
 125          *        IPv6.
 126          * mss_now = pmtu - tp->af_specific->net_header_len -
 127          *           sizeof(struct dccp_hdr) - sizeof(struct dccp_hdr_ext);
 128          */
 129         mss_now = pmtu - sizeof(struct iphdr) - sizeof(struct dccp_hdr) -
 130                   sizeof(struct dccp_hdr_ext);
 131
 132         /* Now subtract optional transport overhead */
 133         mss_now -= dp->dccps_ext_header_len;
 134
 135         /*
 136          * FIXME: this should come from the CCID infrastructure, where, say,
 137          * TFRC will say it wants TIMESTAMPS, ELAPSED time, etc, for now lets
 138          * put a rough estimate for NDP + TIMESTAMP + TIMESTAMP_ECHO + ELAPSED
 139          * TIME + TFRC_OPT_LOSS_EVENT_RATE + TFRC_OPT_RECEIVE_RATE + padding to
 140          * make it a multiple of 4
 141          */
 142
 143         mss_now -= ((5 + 6 + 10 + 6 + 6 + 6 + 3) / 4) * 4;
 144
 145         /* And store cached results */
 146         dp->dccps_pmtu_cookie = pmtu;
 147         dp->dccps_mss_cache = mss_now;
 148
 149         return mss_now;
 150 }
 151
 152 void dccp_write_space(struct sock *sk)
 153 {
 154         read_lock(&sk->sk_callback_lock);
 155
 156         if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
 157                 wake_up_interruptible(sk->sk_sleep);
 158         /* Should agree with poll, otherwise some programs break */
 159         if (sock_writeable(sk))
 160                 sk_wake_async(sk, 2, POLL_OUT);
 161
 162         read_unlock(&sk->sk_callback_lock);
 163 }
 164
 165 /**
 166  * dccp_wait_for_ccid - Wait for ccid to tell us we can send a packet
 167  * @sk: socket to wait for
 168  * @timeo: for how long
 169  */
 170 static int dccp_wait_for_ccid(struct sock *sk, struct sk_buff *skb,
 171                               long *timeo)
 172 {
 173         struct dccp_sock *dp = dccp_sk(sk);
 174         DEFINE_WAIT(wait);
 175         long delay;
 176         int rc;
 177
 178         while (1) {
 179                 prepare_to_wait(sk->sk_sleep, &wait, TASK_INTERRUPTIBLE);
 180
 181                 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 182                         goto do_error;
 183                 if (!*timeo)
 184                         goto do_nonblock;
 185                 if (signal_pending(current))
 186                         goto do_interrupted;
 187
 188                 rc = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
 189                                             skb->len);
 190                 if (rc <= 0)
 191                         break;
 192                 delay = msecs_to_jiffies(rc);
 193                 if (delay > *timeo || delay < 0)
 194                         goto do_nonblock;
 195
 196                 sk->sk_write_pending++;
 197                 release_sock(sk);
 198                 *timeo -= schedule_timeout(delay);
 199                 lock_sock(sk);
 200                 sk->sk_write_pending--;
 201         }
 202 out:
 203         finish_wait(sk->sk_sleep, &wait);
 204         return rc;
 205
 206 do_error:
 207         rc = -EPIPE;
 208         goto out;
 209 do_nonblock:
 210         rc = -EAGAIN;
 211         goto out;
 212 do_interrupted:
 213         rc = sock_intr_errno(*timeo);
 214         goto out;
 215 }
 216
 217 int dccp_write_xmit(struct sock *sk, struct sk_buff *skb, long *timeo)
 218 {
 219         const struct dccp_sock *dp = dccp_sk(sk);
 220         int err = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
 221                                          skb->len);
 222
 223         if (err > 0)
 224                 err = dccp_wait_for_ccid(sk, skb, timeo);
 225
 226         if (err == 0) {
 227                 struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
 228                 const int len = skb->len;
 229
 230                 if (sk->sk_state == DCCP_PARTOPEN) {
 231                         /* See 8.1.5.  Handshake Completion */
 232                         inet_csk_schedule_ack(sk);
 233                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
 234                                                   inet_csk(sk)->icsk_rto,
 235                                                   DCCP_RTO_MAX);
 236                         dcb->dccpd_type = DCCP_PKT_DATAACK;
 237                 } else if (dccp_ack_pending(sk))
 238                         dcb->dccpd_type = DCCP_PKT_DATAACK;
 239                 else
 240                         dcb->dccpd_type = DCCP_PKT_DATA;
 241
 242                 err = dccp_transmit_skb(sk, skb);
 243                 ccid_hc_tx_packet_sent(dp->dccps_hc_tx_ccid, sk, 0, len);
 244         } else
 245                 kfree_skb(skb);
 246
 247         return err;
 248 }
 249
 250 int dccp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
 251 {
 252         if (inet_sk_rebuild_header(sk) != 0)
 253                 return -EHOSTUNREACH; /* Routing failure or similar. */
 254
 255         return dccp_transmit_skb(sk, (skb_cloned(skb) ?
 256                                       pskb_copy(skb, GFP_ATOMIC):
 257                                       skb_clone(skb, GFP_ATOMIC)));
 258 }
 259
 260 struct sk_buff *dccp_make_response(struct sock *sk, struct dst_entry *dst,
 261                                    struct request_sock *req)
 262 {
 263         struct dccp_hdr *dh;
 264         struct dccp_request_sock *dreq;
 265         const int dccp_header_size = sizeof(struct dccp_hdr) +
 266                                      sizeof(struct dccp_hdr_ext) +
 267                                      sizeof(struct dccp_hdr_response);
 268         struct sk_buff *skb = sock_wmalloc(sk, MAX_HEADER + DCCP_MAX_OPT_LEN +
 269                                                dccp_header_size, 1,
 270                                            GFP_ATOMIC);
 271         if (skb == NULL)
 272                 return NULL;
 273
 274         /* Reserve space for headers. */
 275         skb_reserve(skb, MAX_HEADER + DCCP_MAX_OPT_LEN + dccp_header_size);
 276
 277         skb->dst = dst_clone(dst);
 278         skb->csum = 0;
 279
 280         dreq = dccp_rsk(req);
 281         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_RESPONSE;
 282         DCCP_SKB_CB(skb)->dccpd_seq  = dreq->dreq_iss;
 283         dccp_insert_options(sk, skb);
 284
 285         skb->h.raw = skb_push(skb, dccp_header_size);
 286
 287         dh = dccp_hdr(skb);
 288         memset(dh, 0, dccp_header_size);
 289
 290         dh->dccph_sport = inet_sk(sk)->sport;
 291         dh->dccph_dport = inet_rsk(req)->rmt_port;
 292         dh->dccph_doff  = (dccp_header_size +
 293                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
 294         dh->dccph_type  = DCCP_PKT_RESPONSE;
 295         dh->dccph_x     = 1;
 296         dccp_hdr_set_seq(dh, dreq->dreq_iss);
 297         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dreq->dreq_isr);
 298         dccp_hdr_response(skb)->dccph_resp_service = dreq->dreq_service;
 299
 300         dh->dccph_checksum = dccp_v4_checksum(skb, inet_rsk(req)->loc_addr,
 301                                               inet_rsk(req)->rmt_addr);
 302
 303         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
 304         return skb;
 305 }
 306
 307 struct sk_buff *dccp_make_reset(struct sock *sk, struct dst_entry *dst,
 308                                 const enum dccp_reset_codes code)
 309
 310 {
 311         struct dccp_hdr *dh;
 312         struct dccp_sock *dp = dccp_sk(sk);
 313         const int dccp_header_size = sizeof(struct dccp_hdr) +
 314                                      sizeof(struct dccp_hdr_ext) +
 315                                      sizeof(struct dccp_hdr_reset);
 316         struct sk_buff *skb = sock_wmalloc(sk, MAX_HEADER + DCCP_MAX_OPT_LEN +
 317                                                dccp_header_size, 1,
 318                                            GFP_ATOMIC);
 319         if (skb == NULL)
 320                 return NULL;
 321
 322         /* Reserve space for headers. */
 323         skb_reserve(skb, MAX_HEADER + DCCP_MAX_OPT_LEN + dccp_header_size);
 324
 325         skb->dst = dst_clone(dst);
 326         skb->csum = 0;
 327
 328         dccp_inc_seqno(&dp->dccps_gss);
 329
 330         DCCP_SKB_CB(skb)->dccpd_reset_code = code;
 331         DCCP_SKB_CB(skb)->dccpd_type       = DCCP_PKT_RESET;
 332         DCCP_SKB_CB(skb)->dccpd_seq        = dp->dccps_gss;
 333         dccp_insert_options(sk, skb);
 334
 335         skb->h.raw = skb_push(skb, dccp_header_size);
 336
 337         dh = dccp_hdr(skb);
 338         memset(dh, 0, dccp_header_size);
 339
 340         dh->dccph_sport = inet_sk(sk)->sport;
 341         dh->dccph_dport = inet_sk(sk)->dport;
 342         dh->dccph_doff  = (dccp_header_size +
 343                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
 344         dh->dccph_type  = DCCP_PKT_RESET;
 345         dh->dccph_x     = 1;
 346         dccp_hdr_set_seq(dh, dp->dccps_gss);
 347         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dp->dccps_gsr);
 348
 349         dccp_hdr_reset(skb)->dccph_reset_code = code;
 350
 351         dh->dccph_checksum = dccp_v4_checksum(skb, inet_sk(sk)->saddr,
 352                                               inet_sk(sk)->daddr);
 353
 354         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
 355         return skb;
 356 }
 357
 358 /*
 359  * Do all connect socket setups that can be done AF independent.
 360  */
 361 static inline void dccp_connect_init(struct sock *sk)
 362 {
 363         struct dst_entry *dst = __sk_dst_get(sk);
 364         struct inet_connection_sock *icsk = inet_csk(sk);
 365
 366         sk->sk_err = 0;
 367         sock_reset_flag(sk, SOCK_DONE);
 368
 369         dccp_sync_mss(sk, dst_mtu(dst));
 370
 371         /*
 372          * FIXME: set dp->{dccps_swh,dccps_swl}, with
 373          * something like dccp_inc_seq
 374          */
 375
 376         icsk->icsk_retransmits = 0;
 377 }
 378
 379 int dccp_connect(struct sock *sk)
 380 {
 381         struct sk_buff *skb;
 382         struct inet_connection_sock *icsk = inet_csk(sk);
 383
 384         dccp_connect_init(sk);
 385
 386         skb = alloc_skb(MAX_DCCP_HEADER + 15, sk->sk_allocation);
 387         if (unlikely(skb == NULL))
 388                 return -ENOBUFS;
 389
 390         /* Reserve space for headers. */
 391         skb_reserve(skb, MAX_DCCP_HEADER);
 392
 393         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_REQUEST;
 394         skb->csum = 0;
 395         skb_set_owner_w(skb, sk);
 396
 397         BUG_TRAP(sk->sk_send_head == NULL);
 398         sk->sk_send_head = skb;
 399         dccp_transmit_skb(sk, skb_clone(skb, GFP_KERNEL));
 400         DCCP_INC_STATS(DCCP_MIB_ACTIVEOPENS);
 401
 402         /* Timer for repeating the REQUEST until an answer. */
 403         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 404                                   icsk->icsk_rto, DCCP_RTO_MAX);
 405         return 0;
 406 }
 407
 408 void dccp_send_ack(struct sock *sk)
 409 {
 410         /* If we have been reset, we may not send again. */
 411         if (sk->sk_state != DCCP_CLOSED) {
 412                 struct sk_buff *skb = alloc_skb(MAX_DCCP_HEADER, GFP_ATOMIC);
 413
 414                 if (skb == NULL) {
 415                         inet_csk_schedule_ack(sk);
 416                         inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
 417                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
 418                                                   TCP_DELACK_MAX,
 419                                                   DCCP_RTO_MAX);
 420                         return;
 421                 }
 422
 423                 /* Reserve space for headers */
 424                 skb_reserve(skb, MAX_DCCP_HEADER);
 425                 skb->csum = 0;
 426                 DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_ACK;
 427                 skb_set_owner_w(skb, sk);
 428                 dccp_transmit_skb(sk, skb);
 429         }
 430 }
 431
 432 EXPORT_SYMBOL_GPL(dccp_send_ack);
 433
 434 void dccp_send_delayed_ack(struct sock *sk)
 435 {
 436         struct inet_connection_sock *icsk = inet_csk(sk);
 437         /*
 438          * FIXME: tune this timer. elapsed time fixes the skew, so no problem
 439          * with using 2s, and active senders also piggyback the ACK into a
 440          * DATAACK packet, so this is really for quiescent senders.
 441          */
 442         unsigned long timeout = jiffies + 2 * HZ;
 443
 444         /* Use new timeout only if there wasn't a older one earlier. */
 445         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
 446                 /* If delack timer was blocked or is about to expire,
 447                  * send ACK now.
 448                  *
 449                  * FIXME: check the "about to expire" part
 450                  */
 451                 if (icsk->icsk_ack.blocked) {
 452                         dccp_send_ack(sk);
 453                         return;
 454                 }
 455
 456                 if (!time_before(timeout, icsk->icsk_ack.timeout))
 457                         timeout = icsk->icsk_ack.timeout;
 458         }
 459         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
 460         icsk->icsk_ack.timeout = timeout;
 461         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
 462 }
 463
 464 void dccp_send_sync(struct sock *sk, const u64 seq,
 465                     const enum dccp_pkt_type pkt_type)
 466 {
 467         /*
 468          * We are not putting this on the write queue, so
 469          * dccp_transmit_skb() will set the ownership to this
 470          * sock.
 471          */
 472         struct sk_buff *skb = alloc_skb(MAX_DCCP_HEADER, GFP_ATOMIC);
 473
 474         if (skb == NULL)
 475                 /* FIXME: how to make sure the sync is sent? */
 476                 return;
 477
 478         /* Reserve space for headers and prepare control bits. */
 479         skb_reserve(skb, MAX_DCCP_HEADER);
 480         skb->csum = 0;
 481         DCCP_SKB_CB(skb)->dccpd_type = pkt_type;
 482         DCCP_SKB_CB(skb)->dccpd_seq = seq;
 483
 484         skb_set_owner_w(skb, sk);
 485         dccp_transmit_skb(sk, skb);
 486 }
 487
 488 /*
 489  * Send a DCCP_PKT_CLOSE/CLOSEREQ. The caller locks the socket for us. This
 490  * cannot be allowed to fail queueing a DCCP_PKT_CLOSE/CLOSEREQ frame under
 491  * any circumstances.
 492  */
 493 void dccp_send_close(struct sock *sk, const int active)
 494 {
 495         struct dccp_sock *dp = dccp_sk(sk);
 496         struct sk_buff *skb;
 497         const unsigned int prio = active ? GFP_KERNEL : GFP_ATOMIC;
 498
 499         skb = alloc_skb(sk->sk_prot->max_header, prio);
 500         if (skb == NULL)
 501                 return;
 502
 503         /* Reserve space for headers and prepare control bits. */
 504         skb_reserve(skb, sk->sk_prot->max_header);
 505         skb->csum = 0;
 506         DCCP_SKB_CB(skb)->dccpd_type = dp->dccps_role == DCCP_ROLE_CLIENT ?
 507                                         DCCP_PKT_CLOSE : DCCP_PKT_CLOSEREQ;
 508
 509         skb_set_owner_w(skb, sk);
 510         if (active) {
 511                 BUG_TRAP(sk->sk_send_head == NULL);
 512                 sk->sk_send_head = skb;
 513                 dccp_transmit_skb(sk, skb_clone(skb, prio));
 514         } else
 515                 dccp_transmit_skb(sk, skb);
 516 }