net/rxrpc/output.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* RxRPC packet transmission
   3  *
   4  * Copyright (C) 2007 Red Hat, Inc. All Rights Reserved.
   5  * Written by David Howells (dhowells@redhat.com)
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/net.h>
  11 #include <linux/gfp.h>
  12 #include <linux/skbuff.h>
  13 #include <linux/export.h>
  14 #include <net/sock.h>
  15 #include <net/af_rxrpc.h>
  16 #include <net/udp.h>
  17 #include "ar-internal.h"
  18
  19 extern int udpv6_sendmsg(struct sock *sk, struct msghdr *msg, size_t len);
  20
  21 static ssize_t do_udp_sendmsg(struct socket *socket, struct msghdr *msg, size_t len)
  22 {
  23         struct sockaddr *sa = msg->msg_name;
  24         struct sock *sk = socket->sk;
  25
  26         if (IS_ENABLED(CONFIG_AF_RXRPC_IPV6)) {
  27                 if (sa->sa_family == AF_INET6) {
  28                         if (sk->sk_family != AF_INET6) {
  29                                 pr_warn("AF_INET6 address on AF_INET socket\n");
  30                                 return -ENOPROTOOPT;
  31                         }
  32                         return udpv6_sendmsg(sk, msg, len);
  33                 }
  34         }
  35         return udp_sendmsg(sk, msg, len);
  36 }
  37
  38 struct rxrpc_abort_buffer {
  39         struct rxrpc_wire_header whdr;
  40         __be32 abort_code;
  41 };
  42
  43 static const char rxrpc_keepalive_string[] = "";
  44
  45 /*
  46  * Increase Tx backoff on transmission failure and clear it on success.
  47  */
  48 static void rxrpc_tx_backoff(struct rxrpc_call *call, int ret)
  49 {
  50         if (ret < 0) {
  51                 u16 tx_backoff = READ_ONCE(call->tx_backoff);
  52
  53                 if (tx_backoff < HZ)
  54                         WRITE_ONCE(call->tx_backoff, tx_backoff + 1);
  55         } else {
  56                 WRITE_ONCE(call->tx_backoff, 0);
  57         }
  58 }
  59
  60 /*
  61  * Arrange for a keepalive ping a certain time after we last transmitted.  This
  62  * lets the far side know we're still interested in this call and helps keep
  63  * the route through any intervening firewall open.
  64  *
  65  * Receiving a response to the ping will prevent the ->expect_rx_by timer from
  66  * expiring.
  67  */
  68 static void rxrpc_set_keepalive(struct rxrpc_call *call)
  69 {
  70         unsigned long now = jiffies, keepalive_at = call->next_rx_timo / 6;
  71
  72         keepalive_at += now;
  73         WRITE_ONCE(call->keepalive_at, keepalive_at);
  74         rxrpc_reduce_call_timer(call, keepalive_at, now,
  75                                 rxrpc_timer_set_for_keepalive);
  76 }
  77
  78 /*
  79  * Fill out an ACK packet.
  80  */
  81 static size_t rxrpc_fill_out_ack(struct rxrpc_connection *conn,
  82                                  struct rxrpc_call *call,
  83                                  struct rxrpc_txbuf *txb)
  84 {
  85         struct rxrpc_ackinfo ackinfo;
  86         unsigned int qsize;
  87         rxrpc_seq_t window, wtop, wrap_point, ix, first;
  88         int rsize;
  89         u64 wtmp;
  90         u32 mtu, jmax;
  91         u8 *ackp = txb->acks;
  92         u8 sack_buffer[sizeof(call->ackr_sack_table)] __aligned(8);
  93
  94         atomic_set(&call->ackr_nr_unacked, 0);
  95         atomic_set(&call->ackr_nr_consumed, 0);
  96         rxrpc_inc_stat(call->rxnet, stat_tx_ack_fill);
  97
  98         /* Barrier against rxrpc_input_data(). */
  99 retry:
 100         wtmp   = atomic64_read_acquire(&call->ackr_window);
 101         window = lower_32_bits(wtmp);
 102         wtop   = upper_32_bits(wtmp);
 103         txb->ack.firstPacket = htonl(window);
 104         txb->ack.nAcks = 0;
 105
 106         if (after(wtop, window)) {
 107                 /* Try to copy the SACK ring locklessly.  We can use the copy,
 108                  * only if the now-current top of the window didn't go past the
 109                  * previously read base - otherwise we can't know whether we
 110                  * have old data or new data.
 111                  */
 112                 memcpy(sack_buffer, call->ackr_sack_table, sizeof(sack_buffer));
 113                 wrap_point = window + RXRPC_SACK_SIZE - 1;
 114                 wtmp   = atomic64_read_acquire(&call->ackr_window);
 115                 window = lower_32_bits(wtmp);
 116                 wtop   = upper_32_bits(wtmp);
 117                 if (after(wtop, wrap_point)) {
 118                         cond_resched();
 119                         goto retry;
 120                 }
 121
 122                 /* The buffer is maintained as a ring with an invariant mapping
 123                  * between bit position and sequence number, so we'll probably
 124                  * need to rotate it.
 125                  */
 126                 txb->ack.nAcks = wtop - window;
 127                 ix = window % RXRPC_SACK_SIZE;
 128                 first = sizeof(sack_buffer) - ix;
 129
 130                 if (ix + txb->ack.nAcks <= RXRPC_SACK_SIZE) {
 131                         memcpy(txb->acks, sack_buffer + ix, txb->ack.nAcks);
 132                 } else {
 133                         memcpy(txb->acks, sack_buffer + ix, first);
 134                         memcpy(txb->acks + first, sack_buffer,
 135                                txb->ack.nAcks - first);
 136                 }
 137
 138                 ackp += txb->ack.nAcks;
 139         } else if (before(wtop, window)) {
 140                 pr_warn("ack window backward %x %x", window, wtop);
 141         } else if (txb->ack.reason == RXRPC_ACK_DELAY) {
 142                 txb->ack.reason = RXRPC_ACK_IDLE;
 143         }
 144
 145         mtu = conn->peer->if_mtu;
 146         mtu -= conn->peer->hdrsize;
 147         jmax = rxrpc_rx_jumbo_max;
 148         qsize = (window - 1) - call->rx_consumed;
 149         rsize = max_t(int, call->rx_winsize - qsize, 0);
 150         ackinfo.rxMTU           = htonl(rxrpc_rx_mtu);
 151         ackinfo.maxMTU          = htonl(mtu);
 152         ackinfo.rwind           = htonl(rsize);
 153         ackinfo.jumbo_max       = htonl(jmax);
 154
 155         *ackp++ = 0;
 156         *ackp++ = 0;
 157         *ackp++ = 0;
 158         memcpy(ackp, &ackinfo, sizeof(ackinfo));
 159         return txb->ack.nAcks + 3 + sizeof(ackinfo);
 160 }
 161
 162 /*
 163  * Record the beginning of an RTT probe.
 164  */
 165 static int rxrpc_begin_rtt_probe(struct rxrpc_call *call, rxrpc_serial_t serial,
 166                                  enum rxrpc_rtt_tx_trace why)
 167 {
 168         unsigned long avail = call->rtt_avail;
 169         int rtt_slot = 9;
 170
 171         if (!(avail & RXRPC_CALL_RTT_AVAIL_MASK))
 172                 goto no_slot;
 173
 174         rtt_slot = __ffs(avail & RXRPC_CALL_RTT_AVAIL_MASK);
 175         if (!test_and_clear_bit(rtt_slot, &call->rtt_avail))
 176                 goto no_slot;
 177
 178         call->rtt_serial[rtt_slot] = serial;
 179         call->rtt_sent_at[rtt_slot] = ktime_get_real();
 180         smp_wmb(); /* Write data before avail bit */
 181         set_bit(rtt_slot + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
 182
 183         trace_rxrpc_rtt_tx(call, why, rtt_slot, serial);
 184         return rtt_slot;
 185
 186 no_slot:
 187         trace_rxrpc_rtt_tx(call, rxrpc_rtt_tx_no_slot, rtt_slot, serial);
 188         return -1;
 189 }
 190
 191 /*
 192  * Cancel an RTT probe.
 193  */
 194 static void rxrpc_cancel_rtt_probe(struct rxrpc_call *call,
 195                                    rxrpc_serial_t serial, int rtt_slot)
 196 {
 197         if (rtt_slot != -1) {
 198                 clear_bit(rtt_slot + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
 199                 smp_wmb(); /* Clear pending bit before setting slot */
 200                 set_bit(rtt_slot, &call->rtt_avail);
 201                 trace_rxrpc_rtt_tx(call, rxrpc_rtt_tx_cancel, rtt_slot, serial);
 202         }
 203 }
 204
 205 /*
 206  * Send an ACK call packet.
 207  */
 208 static int rxrpc_send_ack_packet(struct rxrpc_local *local, struct rxrpc_txbuf *txb)
 209 {
 210         struct rxrpc_connection *conn;
 211         struct rxrpc_call *call = txb->call;
 212         struct msghdr msg;
 213         struct kvec iov[1];
 214         rxrpc_serial_t serial;
 215         size_t len, n;
 216         int ret, rtt_slot = -1;
 217
 218         if (test_bit(RXRPC_CALL_DISCONNECTED, &call->flags))
 219                 return -ECONNRESET;
 220
 221         conn = call->conn;
 222
 223         msg.msg_name    = &call->peer->srx.transport;
 224         msg.msg_namelen = call->peer->srx.transport_len;
 225         msg.msg_control = NULL;
 226         msg.msg_controllen = 0;
 227         msg.msg_flags   = 0;
 228
 229         if (txb->ack.reason == RXRPC_ACK_PING)
 230                 txb->wire.flags |= RXRPC_REQUEST_ACK;
 231
 232         n = rxrpc_fill_out_ack(conn, call, txb);
 233         if (n == 0)
 234                 return 0;
 235
 236         iov[0].iov_base = &txb->wire;
 237         iov[0].iov_len  = sizeof(txb->wire) + sizeof(txb->ack) + n;
 238         len = iov[0].iov_len;
 239
 240         serial = atomic_inc_return(&conn->serial);
 241         txb->wire.serial = htonl(serial);
 242         trace_rxrpc_tx_ack(call->debug_id, serial,
 243                            ntohl(txb->ack.firstPacket),
 244                            ntohl(txb->ack.serial), txb->ack.reason, txb->ack.nAcks);
 245
 246         if (txb->ack.reason == RXRPC_ACK_PING)
 247                 rtt_slot = rxrpc_begin_rtt_probe(call, serial, rxrpc_rtt_tx_ping);
 248
 249         rxrpc_inc_stat(call->rxnet, stat_tx_ack_send);
 250
 251         /* Grab the highest received seq as late as possible */
 252         txb->ack.previousPacket = htonl(call->rx_highest_seq);
 253
 254         iov_iter_kvec(&msg.msg_iter, WRITE, iov, 1, len);
 255         ret = do_udp_sendmsg(conn->local->socket, &msg, len);
 256         call->peer->last_tx_at = ktime_get_seconds();
 257         if (ret < 0)
 258                 trace_rxrpc_tx_fail(call->debug_id, serial, ret,
 259                                     rxrpc_tx_point_call_ack);
 260         else
 261                 trace_rxrpc_tx_packet(call->debug_id, &txb->wire,
 262                                       rxrpc_tx_point_call_ack);
 263         rxrpc_tx_backoff(call, ret);
 264
 265         if (call->state < RXRPC_CALL_COMPLETE) {
 266                 if (ret < 0)
 267                         rxrpc_cancel_rtt_probe(call, serial, rtt_slot);
 268                 rxrpc_set_keepalive(call);
 269         }
 270
 271         return ret;
 272 }
 273
 274 /*
 275  * ACK transmitter for a local endpoint.  The UDP socket locks around each
 276  * transmission, so we can only transmit one packet at a time, ACK, DATA or
 277  * otherwise.
 278  */
 279 void rxrpc_transmit_ack_packets(struct rxrpc_local *local)
 280 {
 281         LIST_HEAD(queue);
 282         int ret;
 283
 284         rxrpc_see_local(local, rxrpc_local_see_tx_ack);
 285
 286         if (list_empty(&local->ack_tx_queue))
 287                 return;
 288
 289         spin_lock_bh(&local->ack_tx_lock);
 290         list_splice_tail_init(&local->ack_tx_queue, &queue);
 291         spin_unlock_bh(&local->ack_tx_lock);
 292
 293         while (!list_empty(&queue)) {
 294                 struct rxrpc_txbuf *txb =
 295                         list_entry(queue.next, struct rxrpc_txbuf, tx_link);
 296
 297                 ret = rxrpc_send_ack_packet(local, txb);
 298                 if (ret < 0 && ret != -ECONNRESET) {
 299                         spin_lock_bh(&local->ack_tx_lock);
 300                         list_splice_init(&queue, &local->ack_tx_queue);
 301                         spin_unlock_bh(&local->ack_tx_lock);
 302                         break;
 303                 }
 304
 305                 list_del_init(&txb->tx_link);
 306                 rxrpc_put_call(txb->call, rxrpc_call_put_send_ack);
 307                 rxrpc_put_txbuf(txb, rxrpc_txbuf_put_ack_tx);
 308         }
 309 }
 310
 311 /*
 312  * Send an ABORT call packet.
 313  */
 314 int rxrpc_send_abort_packet(struct rxrpc_call *call)
 315 {
 316         struct rxrpc_connection *conn;
 317         struct rxrpc_abort_buffer pkt;
 318         struct msghdr msg;
 319         struct kvec iov[1];
 320         rxrpc_serial_t serial;
 321         int ret;
 322
 323         /* Don't bother sending aborts for a client call once the server has
 324          * hard-ACK'd all of its request data.  After that point, we're not
 325          * going to stop the operation proceeding, and whilst we might limit
 326          * the reply, it's not worth it if we can send a new call on the same
 327          * channel instead, thereby closing off this call.
 328          */
 329         if (rxrpc_is_client_call(call) &&
 330             test_bit(RXRPC_CALL_TX_ALL_ACKED, &call->flags))
 331                 return 0;
 332
 333         if (test_bit(RXRPC_CALL_DISCONNECTED, &call->flags))
 334                 return -ECONNRESET;
 335
 336         conn = call->conn;
 337
 338         msg.msg_name    = &call->peer->srx.transport;
 339         msg.msg_namelen = call->peer->srx.transport_len;
 340         msg.msg_control = NULL;
 341         msg.msg_controllen = 0;
 342         msg.msg_flags   = 0;
 343
 344         pkt.whdr.epoch          = htonl(conn->proto.epoch);
 345         pkt.whdr.cid            = htonl(call->cid);
 346         pkt.whdr.callNumber     = htonl(call->call_id);
 347         pkt.whdr.seq            = 0;
 348         pkt.whdr.type           = RXRPC_PACKET_TYPE_ABORT;
 349         pkt.whdr.flags          = conn->out_clientflag;
 350         pkt.whdr.userStatus     = 0;
 351         pkt.whdr.securityIndex  = call->security_ix;
 352         pkt.whdr._rsvd          = 0;
 353         pkt.whdr.serviceId      = htons(call->dest_srx.srx_service);
 354         pkt.abort_code          = htonl(call->abort_code);
 355
 356         iov[0].iov_base = &pkt;
 357         iov[0].iov_len  = sizeof(pkt);
 358
 359         serial = atomic_inc_return(&conn->serial);
 360         pkt.whdr.serial = htonl(serial);
 361
 362         iov_iter_kvec(&msg.msg_iter, WRITE, iov, 1, sizeof(pkt));
 363         ret = do_udp_sendmsg(conn->local->socket, &msg, sizeof(pkt));
 364         conn->peer->last_tx_at = ktime_get_seconds();
 365         if (ret < 0)
 366                 trace_rxrpc_tx_fail(call->debug_id, serial, ret,
 367                                     rxrpc_tx_point_call_abort);
 368         else
 369                 trace_rxrpc_tx_packet(call->debug_id, &pkt.whdr,
 370                                       rxrpc_tx_point_call_abort);
 371         rxrpc_tx_backoff(call, ret);
 372         return ret;
 373 }
 374
 375 /*
 376  * send a packet through the transport endpoint
 377  */
 378 int rxrpc_send_data_packet(struct rxrpc_call *call, struct rxrpc_txbuf *txb)
 379 {
 380         enum rxrpc_req_ack_trace why;
 381         struct rxrpc_connection *conn = call->conn;
 382         struct msghdr msg;
 383         struct kvec iov[1];
 384         rxrpc_serial_t serial;
 385         size_t len;
 386         int ret, rtt_slot = -1;
 387
 388         _enter("%x,{%d}", txb->seq, txb->len);
 389
 390         /* Each transmission of a Tx packet needs a new serial number */
 391         serial = atomic_inc_return(&conn->serial);
 392         txb->wire.serial = htonl(serial);
 393
 394         if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags) &&
 395             txb->seq == 1)
 396                 txb->wire.userStatus = RXRPC_USERSTATUS_SERVICE_UPGRADE;
 397
 398         iov[0].iov_base = &txb->wire;
 399         iov[0].iov_len = sizeof(txb->wire) + txb->len;
 400         len = iov[0].iov_len;
 401         iov_iter_kvec(&msg.msg_iter, WRITE, iov, 1, len);
 402
 403         msg.msg_name = &call->peer->srx.transport;
 404         msg.msg_namelen = call->peer->srx.transport_len;
 405         msg.msg_control = NULL;
 406         msg.msg_controllen = 0;
 407         msg.msg_flags = 0;
 408
 409         /* If our RTT cache needs working on, request an ACK.  Also request
 410          * ACKs if a DATA packet appears to have been lost.
 411          *
 412          * However, we mustn't request an ACK on the last reply packet of a
 413          * service call, lest OpenAFS incorrectly send us an ACK with some
 414          * soft-ACKs in it and then never follow up with a proper hard ACK.
 415          */
 416         if (txb->wire.flags & RXRPC_REQUEST_ACK)
 417                 why = rxrpc_reqack_already_on;
 418         else if (test_bit(RXRPC_TXBUF_LAST, &txb->flags) && rxrpc_sending_to_client(txb))
 419                 why = rxrpc_reqack_no_srv_last;
 420         else if (test_and_clear_bit(RXRPC_CALL_EV_ACK_LOST, &call->events))
 421                 why = rxrpc_reqack_ack_lost;
 422         else if (test_bit(RXRPC_TXBUF_RESENT, &txb->flags))
 423                 why = rxrpc_reqack_retrans;
 424         else if (call->cong_mode == RXRPC_CALL_SLOW_START && call->cong_cwnd <= 2)
 425                 why = rxrpc_reqack_slow_start;
 426         else if (call->tx_winsize <= 2)
 427                 why = rxrpc_reqack_small_txwin;
 428         else if (call->peer->rtt_count < 3 && txb->seq & 1)
 429                 why = rxrpc_reqack_more_rtt;
 430         else if (ktime_before(ktime_add_ms(call->peer->rtt_last_req, 1000), ktime_get_real()))
 431                 why = rxrpc_reqack_old_rtt;
 432         else
 433                 goto dont_set_request_ack;
 434
 435         rxrpc_inc_stat(call->rxnet, stat_why_req_ack[why]);
 436         trace_rxrpc_req_ack(call->debug_id, txb->seq, why);
 437         if (why != rxrpc_reqack_no_srv_last)
 438                 txb->wire.flags |= RXRPC_REQUEST_ACK;
 439 dont_set_request_ack:
 440
 441         if (IS_ENABLED(CONFIG_AF_RXRPC_INJECT_LOSS)) {
 442                 static int lose;
 443                 if ((lose++ & 7) == 7) {
 444                         ret = 0;
 445                         trace_rxrpc_tx_data(call, txb->seq, serial,
 446                                             txb->wire.flags,
 447                                             test_bit(RXRPC_TXBUF_RESENT, &txb->flags),
 448                                             true);
 449                         goto done;
 450                 }
 451         }
 452
 453         trace_rxrpc_tx_data(call, txb->seq, serial, txb->wire.flags,
 454                             test_bit(RXRPC_TXBUF_RESENT, &txb->flags), false);
 455
 456         /* Track what we've attempted to transmit at least once so that the
 457          * retransmission algorithm doesn't try to resend what we haven't sent
 458          * yet.  However, this can race as we can receive an ACK before we get
 459          * to this point.  But, OTOH, if we won't get an ACK mentioning this
 460          * packet unless the far side received it (though it could have
 461          * discarded it anyway and NAK'd it).
 462          */
 463         cmpxchg(&call->tx_transmitted, txb->seq - 1, txb->seq);
 464
 465         /* send the packet with the don't fragment bit set if we currently
 466          * think it's small enough */
 467         if (txb->len >= call->peer->maxdata)
 468                 goto send_fragmentable;
 469
 470         down_read(&conn->local->defrag_sem);
 471
 472         txb->last_sent = ktime_get_real();
 473         if (txb->wire.flags & RXRPC_REQUEST_ACK)
 474                 rtt_slot = rxrpc_begin_rtt_probe(call, serial, rxrpc_rtt_tx_data);
 475
 476         /* send the packet by UDP
 477          * - returns -EMSGSIZE if UDP would have to fragment the packet
 478          *   to go out of the interface
 479          *   - in which case, we'll have processed the ICMP error
 480          *     message and update the peer record
 481          */
 482         rxrpc_inc_stat(call->rxnet, stat_tx_data_send);
 483         ret = do_udp_sendmsg(conn->local->socket, &msg, len);
 484         conn->peer->last_tx_at = ktime_get_seconds();
 485
 486         up_read(&conn->local->defrag_sem);
 487         if (ret < 0) {
 488                 rxrpc_cancel_rtt_probe(call, serial, rtt_slot);
 489                 trace_rxrpc_tx_fail(call->debug_id, serial, ret,
 490                                     rxrpc_tx_point_call_data_nofrag);
 491         } else {
 492                 trace_rxrpc_tx_packet(call->debug_id, &txb->wire,
 493                                       rxrpc_tx_point_call_data_nofrag);
 494         }
 495
 496         rxrpc_tx_backoff(call, ret);
 497         if (ret == -EMSGSIZE)
 498                 goto send_fragmentable;
 499
 500 done:
 501         if (ret >= 0) {
 502                 call->tx_last_sent = txb->last_sent;
 503                 if (txb->wire.flags & RXRPC_REQUEST_ACK) {
 504                         call->peer->rtt_last_req = txb->last_sent;
 505                         if (call->peer->rtt_count > 1) {
 506                                 unsigned long nowj = jiffies, ack_lost_at;
 507
 508                                 ack_lost_at = rxrpc_get_rto_backoff(call->peer, false);
 509                                 ack_lost_at += nowj;
 510                                 WRITE_ONCE(call->ack_lost_at, ack_lost_at);
 511                                 rxrpc_reduce_call_timer(call, ack_lost_at, nowj,
 512                                                         rxrpc_timer_set_for_lost_ack);
 513                         }
 514                 }
 515
 516                 if (txb->seq == 1 &&
 517                     !test_and_set_bit(RXRPC_CALL_BEGAN_RX_TIMER,
 518                                       &call->flags)) {
 519                         unsigned long nowj = jiffies, expect_rx_by;
 520
 521                         expect_rx_by = nowj + call->next_rx_timo;
 522                         WRITE_ONCE(call->expect_rx_by, expect_rx_by);
 523                         rxrpc_reduce_call_timer(call, expect_rx_by, nowj,
 524                                                 rxrpc_timer_set_for_normal);
 525                 }
 526
 527                 rxrpc_set_keepalive(call);
 528         } else {
 529                 /* Cancel the call if the initial transmission fails,
 530                  * particularly if that's due to network routing issues that
 531                  * aren't going away anytime soon.  The layer above can arrange
 532                  * the retransmission.
 533                  */
 534                 if (!test_and_set_bit(RXRPC_CALL_BEGAN_RX_TIMER, &call->flags))
 535                         rxrpc_set_call_completion(call, RXRPC_CALL_LOCAL_ERROR,
 536                                                   RX_USER_ABORT, ret);
 537         }
 538
 539         _leave(" = %d [%u]", ret, call->peer->maxdata);
 540         return ret;
 541
 542 send_fragmentable:
 543         /* attempt to send this message with fragmentation enabled */
 544         _debug("send fragment");
 545
 546         down_write(&conn->local->defrag_sem);
 547
 548         txb->last_sent = ktime_get_real();
 549         if (txb->wire.flags & RXRPC_REQUEST_ACK)
 550                 rtt_slot = rxrpc_begin_rtt_probe(call, serial, rxrpc_rtt_tx_data);
 551
 552         switch (conn->local->srx.transport.family) {
 553         case AF_INET6:
 554         case AF_INET:
 555                 ip_sock_set_mtu_discover(conn->local->socket->sk,
 556                                          IP_PMTUDISC_DONT);
 557                 rxrpc_inc_stat(call->rxnet, stat_tx_data_send_frag);
 558                 ret = do_udp_sendmsg(conn->local->socket, &msg, len);
 559                 conn->peer->last_tx_at = ktime_get_seconds();
 560
 561                 ip_sock_set_mtu_discover(conn->local->socket->sk,
 562                                          IP_PMTUDISC_DO);
 563                 break;
 564
 565         default:
 566                 BUG();
 567         }
 568
 569         if (ret < 0) {
 570                 rxrpc_cancel_rtt_probe(call, serial, rtt_slot);
 571                 trace_rxrpc_tx_fail(call->debug_id, serial, ret,
 572                                     rxrpc_tx_point_call_data_frag);
 573         } else {
 574                 trace_rxrpc_tx_packet(call->debug_id, &txb->wire,
 575                                       rxrpc_tx_point_call_data_frag);
 576         }
 577         rxrpc_tx_backoff(call, ret);
 578
 579         up_write(&conn->local->defrag_sem);
 580         goto done;
 581 }
 582
 583 /*
 584  * Reject a packet through the local endpoint.
 585  */
 586 void rxrpc_reject_packet(struct rxrpc_local *local, struct sk_buff *skb)
 587 {
 588         struct rxrpc_wire_header whdr;
 589         struct sockaddr_rxrpc srx;
 590         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 591         struct msghdr msg;
 592         struct kvec iov[2];
 593         size_t size;
 594         __be32 code;
 595         int ret, ioc;
 596
 597         rxrpc_see_skb(skb, rxrpc_skb_see_reject);
 598
 599         iov[0].iov_base = &whdr;
 600         iov[0].iov_len = sizeof(whdr);
 601         iov[1].iov_base = &code;
 602         iov[1].iov_len = sizeof(code);
 603
 604         msg.msg_name = &srx.transport;
 605         msg.msg_control = NULL;
 606         msg.msg_controllen = 0;
 607         msg.msg_flags = 0;
 608
 609         memset(&whdr, 0, sizeof(whdr));
 610
 611         switch (skb->mark) {
 612         case RXRPC_SKB_MARK_REJECT_BUSY:
 613                 whdr.type = RXRPC_PACKET_TYPE_BUSY;
 614                 size = sizeof(whdr);
 615                 ioc = 1;
 616                 break;
 617         case RXRPC_SKB_MARK_REJECT_ABORT:
 618                 whdr.type = RXRPC_PACKET_TYPE_ABORT;
 619                 code = htonl(skb->priority);
 620                 size = sizeof(whdr) + sizeof(code);
 621                 ioc = 2;
 622                 break;
 623         default:
 624                 return;
 625         }
 626
 627         if (rxrpc_extract_addr_from_skb(&srx, skb) == 0) {
 628                 msg.msg_namelen = srx.transport_len;
 629
 630                 whdr.epoch      = htonl(sp->hdr.epoch);
 631                 whdr.cid        = htonl(sp->hdr.cid);
 632                 whdr.callNumber = htonl(sp->hdr.callNumber);
 633                 whdr.serviceId  = htons(sp->hdr.serviceId);
 634                 whdr.flags      = sp->hdr.flags;
 635                 whdr.flags      ^= RXRPC_CLIENT_INITIATED;
 636                 whdr.flags      &= RXRPC_CLIENT_INITIATED;
 637
 638                 iov_iter_kvec(&msg.msg_iter, WRITE, iov, ioc, size);
 639                 ret = do_udp_sendmsg(local->socket, &msg, size);
 640                 if (ret < 0)
 641                         trace_rxrpc_tx_fail(local->debug_id, 0, ret,
 642                                             rxrpc_tx_point_reject);
 643                 else
 644                         trace_rxrpc_tx_packet(local->debug_id, &whdr,
 645                                               rxrpc_tx_point_reject);
 646         }
 647 }
 648
 649 /*
 650  * Send a VERSION reply to a peer as a keepalive.
 651  */
 652 void rxrpc_send_keepalive(struct rxrpc_peer *peer)
 653 {
 654         struct rxrpc_wire_header whdr;
 655         struct msghdr msg;
 656         struct kvec iov[2];
 657         size_t len;
 658         int ret;
 659
 660         _enter("");
 661
 662         msg.msg_name    = &peer->srx.transport;
 663         msg.msg_namelen = peer->srx.transport_len;
 664         msg.msg_control = NULL;
 665         msg.msg_controllen = 0;
 666         msg.msg_flags   = 0;
 667
 668         whdr.epoch      = htonl(peer->local->rxnet->epoch);
 669         whdr.cid        = 0;
 670         whdr.callNumber = 0;
 671         whdr.seq        = 0;
 672         whdr.serial     = 0;
 673         whdr.type       = RXRPC_PACKET_TYPE_VERSION; /* Not client-initiated */
 674         whdr.flags      = RXRPC_LAST_PACKET;
 675         whdr.userStatus = 0;
 676         whdr.securityIndex = 0;
 677         whdr._rsvd      = 0;
 678         whdr.serviceId  = 0;
 679
 680         iov[0].iov_base = &whdr;
 681         iov[0].iov_len  = sizeof(whdr);
 682         iov[1].iov_base = (char *)rxrpc_keepalive_string;
 683         iov[1].iov_len  = sizeof(rxrpc_keepalive_string);
 684
 685         len = iov[0].iov_len + iov[1].iov_len;
 686
 687         iov_iter_kvec(&msg.msg_iter, WRITE, iov, 2, len);
 688         ret = do_udp_sendmsg(peer->local->socket, &msg, len);
 689         if (ret < 0)
 690                 trace_rxrpc_tx_fail(peer->debug_id, 0, ret,
 691                                     rxrpc_tx_point_version_keepalive);
 692         else
 693                 trace_rxrpc_tx_packet(peer->debug_id, &whdr,
 694                                       rxrpc_tx_point_version_keepalive);
 695
 696         peer->last_tx_at = ktime_get_seconds();
 697         _leave("");
 698 }
 699
 700 /*
 701  * Schedule an instant Tx resend.
 702  */
 703 static inline void rxrpc_instant_resend(struct rxrpc_call *call,
 704                                         struct rxrpc_txbuf *txb)
 705 {
 706         if (call->state < RXRPC_CALL_COMPLETE)
 707                 kdebug("resend");
 708 }
 709
 710 /*
 711  * Transmit one packet.
 712  */
 713 void rxrpc_transmit_one(struct rxrpc_call *call, struct rxrpc_txbuf *txb)
 714 {
 715         int ret;
 716
 717         ret = rxrpc_send_data_packet(call, txb);
 718         if (ret < 0) {
 719                 switch (ret) {
 720                 case -ENETUNREACH:
 721                 case -EHOSTUNREACH:
 722                 case -ECONNREFUSED:
 723                         rxrpc_set_call_completion(call, RXRPC_CALL_LOCAL_ERROR,
 724                                                   0, ret);
 725                         break;
 726                 default:
 727                         _debug("need instant resend %d", ret);
 728                         rxrpc_instant_resend(call, txb);
 729                 }
 730         } else {
 731                 unsigned long now = jiffies;
 732                 unsigned long resend_at = now + call->peer->rto_j;
 733
 734                 WRITE_ONCE(call->resend_at, resend_at);
 735                 rxrpc_reduce_call_timer(call, resend_at, now,
 736                                         rxrpc_timer_set_for_send);
 737         }
 738 }