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  * Transmit an ACK packet.
 207  */
 208 int rxrpc_send_ack_packet(struct rxrpc_call *call, struct rxrpc_txbuf *txb)
 209 {
 210         struct rxrpc_connection *conn;
 211         struct msghdr msg;
 212         struct kvec iov[1];
 213         rxrpc_serial_t serial;
 214         size_t len, n;
 215         int ret, rtt_slot = -1;
 216
 217         if (test_bit(RXRPC_CALL_DISCONNECTED, &call->flags))
 218                 return -ECONNRESET;
 219
 220         conn = call->conn;
 221
 222         msg.msg_name    = &call->peer->srx.transport;
 223         msg.msg_namelen = call->peer->srx.transport_len;
 224         msg.msg_control = NULL;
 225         msg.msg_controllen = 0;
 226         msg.msg_flags   = 0;
 227
 228         if (txb->ack.reason == RXRPC_ACK_PING)
 229                 txb->wire.flags |= RXRPC_REQUEST_ACK;
 230
 231         n = rxrpc_fill_out_ack(conn, call, txb);
 232         if (n == 0)
 233                 return 0;
 234
 235         iov[0].iov_base = &txb->wire;
 236         iov[0].iov_len  = sizeof(txb->wire) + sizeof(txb->ack) + n;
 237         len = iov[0].iov_len;
 238
 239         serial = atomic_inc_return(&conn->serial);
 240         txb->wire.serial = htonl(serial);
 241         trace_rxrpc_tx_ack(call->debug_id, serial,
 242                            ntohl(txb->ack.firstPacket),
 243                            ntohl(txb->ack.serial), txb->ack.reason, txb->ack.nAcks);
 244
 245         if (txb->ack.reason == RXRPC_ACK_PING)
 246                 rtt_slot = rxrpc_begin_rtt_probe(call, serial, rxrpc_rtt_tx_ping);
 247
 248         rxrpc_inc_stat(call->rxnet, stat_tx_ack_send);
 249
 250         /* Grab the highest received seq as late as possible */
 251         txb->ack.previousPacket = htonl(call->rx_highest_seq);
 252
 253         iov_iter_kvec(&msg.msg_iter, WRITE, iov, 1, len);
 254         ret = do_udp_sendmsg(conn->local->socket, &msg, len);
 255         call->peer->last_tx_at = ktime_get_seconds();
 256         if (ret < 0) {
 257                 trace_rxrpc_tx_fail(call->debug_id, serial, ret,
 258                                     rxrpc_tx_point_call_ack);
 259         } else {
 260                 trace_rxrpc_tx_packet(call->debug_id, &txb->wire,
 261                                       rxrpc_tx_point_call_ack);
 262                 if (txb->wire.flags & RXRPC_REQUEST_ACK)
 263                         call->peer->rtt_last_req = ktime_get_real();
 264         }
 265         rxrpc_tx_backoff(call, ret);
 266
 267         if (!__rxrpc_call_is_complete(call)) {
 268                 if (ret < 0)
 269                         rxrpc_cancel_rtt_probe(call, serial, rtt_slot);
 270                 rxrpc_set_keepalive(call);
 271         }
 272
 273         return ret;
 274 }
 275
 276 /*
 277  * Send an ABORT call packet.
 278  */
 279 int rxrpc_send_abort_packet(struct rxrpc_call *call)
 280 {
 281         struct rxrpc_connection *conn;
 282         struct rxrpc_abort_buffer pkt;
 283         struct msghdr msg;
 284         struct kvec iov[1];
 285         rxrpc_serial_t serial;
 286         int ret;
 287
 288         /* Don't bother sending aborts for a client call once the server has
 289          * hard-ACK'd all of its request data.  After that point, we're not
 290          * going to stop the operation proceeding, and whilst we might limit
 291          * the reply, it's not worth it if we can send a new call on the same
 292          * channel instead, thereby closing off this call.
 293          */
 294         if (rxrpc_is_client_call(call) &&
 295             test_bit(RXRPC_CALL_TX_ALL_ACKED, &call->flags))
 296                 return 0;
 297
 298         if (test_bit(RXRPC_CALL_DISCONNECTED, &call->flags))
 299                 return -ECONNRESET;
 300
 301         conn = call->conn;
 302
 303         msg.msg_name    = &call->peer->srx.transport;
 304         msg.msg_namelen = call->peer->srx.transport_len;
 305         msg.msg_control = NULL;
 306         msg.msg_controllen = 0;
 307         msg.msg_flags   = 0;
 308
 309         pkt.whdr.epoch          = htonl(conn->proto.epoch);
 310         pkt.whdr.cid            = htonl(call->cid);
 311         pkt.whdr.callNumber     = htonl(call->call_id);
 312         pkt.whdr.seq            = 0;
 313         pkt.whdr.type           = RXRPC_PACKET_TYPE_ABORT;
 314         pkt.whdr.flags          = conn->out_clientflag;
 315         pkt.whdr.userStatus     = 0;
 316         pkt.whdr.securityIndex  = call->security_ix;
 317         pkt.whdr._rsvd          = 0;
 318         pkt.whdr.serviceId      = htons(call->dest_srx.srx_service);
 319         pkt.abort_code          = htonl(call->abort_code);
 320
 321         iov[0].iov_base = &pkt;
 322         iov[0].iov_len  = sizeof(pkt);
 323
 324         serial = atomic_inc_return(&conn->serial);
 325         pkt.whdr.serial = htonl(serial);
 326
 327         iov_iter_kvec(&msg.msg_iter, WRITE, iov, 1, sizeof(pkt));
 328         ret = do_udp_sendmsg(conn->local->socket, &msg, sizeof(pkt));
 329         conn->peer->last_tx_at = ktime_get_seconds();
 330         if (ret < 0)
 331                 trace_rxrpc_tx_fail(call->debug_id, serial, ret,
 332                                     rxrpc_tx_point_call_abort);
 333         else
 334                 trace_rxrpc_tx_packet(call->debug_id, &pkt.whdr,
 335                                       rxrpc_tx_point_call_abort);
 336         rxrpc_tx_backoff(call, ret);
 337         return ret;
 338 }
 339
 340 /*
 341  * send a packet through the transport endpoint
 342  */
 343 int rxrpc_send_data_packet(struct rxrpc_call *call, struct rxrpc_txbuf *txb)
 344 {
 345         enum rxrpc_req_ack_trace why;
 346         struct rxrpc_connection *conn = call->conn;
 347         struct msghdr msg;
 348         struct kvec iov[1];
 349         rxrpc_serial_t serial;
 350         size_t len;
 351         int ret, rtt_slot = -1;
 352
 353         _enter("%x,{%d}", txb->seq, txb->len);
 354
 355         /* Each transmission of a Tx packet needs a new serial number */
 356         serial = atomic_inc_return(&conn->serial);
 357         txb->wire.serial = htonl(serial);
 358
 359         if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags) &&
 360             txb->seq == 1)
 361                 txb->wire.userStatus = RXRPC_USERSTATUS_SERVICE_UPGRADE;
 362
 363         iov[0].iov_base = &txb->wire;
 364         iov[0].iov_len = sizeof(txb->wire) + txb->len;
 365         len = iov[0].iov_len;
 366         iov_iter_kvec(&msg.msg_iter, WRITE, iov, 1, len);
 367
 368         msg.msg_name = &call->peer->srx.transport;
 369         msg.msg_namelen = call->peer->srx.transport_len;
 370         msg.msg_control = NULL;
 371         msg.msg_controllen = 0;
 372         msg.msg_flags = 0;
 373
 374         /* If our RTT cache needs working on, request an ACK.  Also request
 375          * ACKs if a DATA packet appears to have been lost.
 376          *
 377          * However, we mustn't request an ACK on the last reply packet of a
 378          * service call, lest OpenAFS incorrectly send us an ACK with some
 379          * soft-ACKs in it and then never follow up with a proper hard ACK.
 380          */
 381         if (txb->wire.flags & RXRPC_REQUEST_ACK)
 382                 why = rxrpc_reqack_already_on;
 383         else if (test_bit(RXRPC_TXBUF_LAST, &txb->flags) && rxrpc_sending_to_client(txb))
 384                 why = rxrpc_reqack_no_srv_last;
 385         else if (test_and_clear_bit(RXRPC_CALL_EV_ACK_LOST, &call->events))
 386                 why = rxrpc_reqack_ack_lost;
 387         else if (test_bit(RXRPC_TXBUF_RESENT, &txb->flags))
 388                 why = rxrpc_reqack_retrans;
 389         else if (call->cong_mode == RXRPC_CALL_SLOW_START && call->cong_cwnd <= 2)
 390                 why = rxrpc_reqack_slow_start;
 391         else if (call->tx_winsize <= 2)
 392                 why = rxrpc_reqack_small_txwin;
 393         else if (call->peer->rtt_count < 3 && txb->seq & 1)
 394                 why = rxrpc_reqack_more_rtt;
 395         else if (ktime_before(ktime_add_ms(call->peer->rtt_last_req, 1000), ktime_get_real()))
 396                 why = rxrpc_reqack_old_rtt;
 397         else
 398                 goto dont_set_request_ack;
 399
 400         rxrpc_inc_stat(call->rxnet, stat_why_req_ack[why]);
 401         trace_rxrpc_req_ack(call->debug_id, txb->seq, why);
 402         if (why != rxrpc_reqack_no_srv_last)
 403                 txb->wire.flags |= RXRPC_REQUEST_ACK;
 404 dont_set_request_ack:
 405
 406         if (IS_ENABLED(CONFIG_AF_RXRPC_INJECT_LOSS)) {
 407                 static int lose;
 408                 if ((lose++ & 7) == 7) {
 409                         ret = 0;
 410                         trace_rxrpc_tx_data(call, txb->seq, serial,
 411                                             txb->wire.flags,
 412                                             test_bit(RXRPC_TXBUF_RESENT, &txb->flags),
 413                                             true);
 414                         goto done;
 415                 }
 416         }
 417
 418         trace_rxrpc_tx_data(call, txb->seq, serial, txb->wire.flags,
 419                             test_bit(RXRPC_TXBUF_RESENT, &txb->flags), false);
 420
 421         /* Track what we've attempted to transmit at least once so that the
 422          * retransmission algorithm doesn't try to resend what we haven't sent
 423          * yet.  However, this can race as we can receive an ACK before we get
 424          * to this point.  But, OTOH, if we won't get an ACK mentioning this
 425          * packet unless the far side received it (though it could have
 426          * discarded it anyway and NAK'd it).
 427          */
 428         cmpxchg(&call->tx_transmitted, txb->seq - 1, txb->seq);
 429
 430         /* send the packet with the don't fragment bit set if we currently
 431          * think it's small enough */
 432         if (txb->len >= call->peer->maxdata)
 433                 goto send_fragmentable;
 434
 435         down_read(&conn->local->defrag_sem);
 436
 437         txb->last_sent = ktime_get_real();
 438         if (txb->wire.flags & RXRPC_REQUEST_ACK)
 439                 rtt_slot = rxrpc_begin_rtt_probe(call, serial, rxrpc_rtt_tx_data);
 440
 441         /* send the packet by UDP
 442          * - returns -EMSGSIZE if UDP would have to fragment the packet
 443          *   to go out of the interface
 444          *   - in which case, we'll have processed the ICMP error
 445          *     message and update the peer record
 446          */
 447         rxrpc_inc_stat(call->rxnet, stat_tx_data_send);
 448         ret = do_udp_sendmsg(conn->local->socket, &msg, len);
 449         conn->peer->last_tx_at = ktime_get_seconds();
 450
 451         up_read(&conn->local->defrag_sem);
 452         if (ret < 0) {
 453                 rxrpc_inc_stat(call->rxnet, stat_tx_data_send_fail);
 454                 rxrpc_cancel_rtt_probe(call, serial, rtt_slot);
 455                 trace_rxrpc_tx_fail(call->debug_id, serial, ret,
 456                                     rxrpc_tx_point_call_data_nofrag);
 457         } else {
 458                 trace_rxrpc_tx_packet(call->debug_id, &txb->wire,
 459                                       rxrpc_tx_point_call_data_nofrag);
 460         }
 461
 462         rxrpc_tx_backoff(call, ret);
 463         if (ret == -EMSGSIZE)
 464                 goto send_fragmentable;
 465
 466 done:
 467         if (ret >= 0) {
 468                 call->tx_last_sent = txb->last_sent;
 469                 if (txb->wire.flags & RXRPC_REQUEST_ACK) {
 470                         call->peer->rtt_last_req = txb->last_sent;
 471                         if (call->peer->rtt_count > 1) {
 472                                 unsigned long nowj = jiffies, ack_lost_at;
 473
 474                                 ack_lost_at = rxrpc_get_rto_backoff(call->peer, false);
 475                                 ack_lost_at += nowj;
 476                                 WRITE_ONCE(call->ack_lost_at, ack_lost_at);
 477                                 rxrpc_reduce_call_timer(call, ack_lost_at, nowj,
 478                                                         rxrpc_timer_set_for_lost_ack);
 479                         }
 480                 }
 481
 482                 if (txb->seq == 1 &&
 483                     !test_and_set_bit(RXRPC_CALL_BEGAN_RX_TIMER,
 484                                       &call->flags)) {
 485                         unsigned long nowj = jiffies, expect_rx_by;
 486
 487                         expect_rx_by = nowj + call->next_rx_timo;
 488                         WRITE_ONCE(call->expect_rx_by, expect_rx_by);
 489                         rxrpc_reduce_call_timer(call, expect_rx_by, nowj,
 490                                                 rxrpc_timer_set_for_normal);
 491                 }
 492
 493                 rxrpc_set_keepalive(call);
 494         } else {
 495                 /* Cancel the call if the initial transmission fails,
 496                  * particularly if that's due to network routing issues that
 497                  * aren't going away anytime soon.  The layer above can arrange
 498                  * the retransmission.
 499                  */
 500                 if (!test_and_set_bit(RXRPC_CALL_BEGAN_RX_TIMER, &call->flags))
 501                         rxrpc_set_call_completion(call, RXRPC_CALL_LOCAL_ERROR,
 502                                                   RX_USER_ABORT, ret);
 503         }
 504
 505         _leave(" = %d [%u]", ret, call->peer->maxdata);
 506         return ret;
 507
 508 send_fragmentable:
 509         /* attempt to send this message with fragmentation enabled */
 510         _debug("send fragment");
 511
 512         down_write(&conn->local->defrag_sem);
 513
 514         txb->last_sent = ktime_get_real();
 515         if (txb->wire.flags & RXRPC_REQUEST_ACK)
 516                 rtt_slot = rxrpc_begin_rtt_probe(call, serial, rxrpc_rtt_tx_data);
 517
 518         switch (conn->local->srx.transport.family) {
 519         case AF_INET6:
 520         case AF_INET:
 521                 ip_sock_set_mtu_discover(conn->local->socket->sk,
 522                                          IP_PMTUDISC_DONT);
 523                 rxrpc_inc_stat(call->rxnet, stat_tx_data_send_frag);
 524                 ret = do_udp_sendmsg(conn->local->socket, &msg, len);
 525                 conn->peer->last_tx_at = ktime_get_seconds();
 526
 527                 ip_sock_set_mtu_discover(conn->local->socket->sk,
 528                                          IP_PMTUDISC_DO);
 529                 break;
 530
 531         default:
 532                 BUG();
 533         }
 534
 535         if (ret < 0) {
 536                 rxrpc_inc_stat(call->rxnet, stat_tx_data_send_fail);
 537                 rxrpc_cancel_rtt_probe(call, serial, rtt_slot);
 538                 trace_rxrpc_tx_fail(call->debug_id, serial, ret,
 539                                     rxrpc_tx_point_call_data_frag);
 540         } else {
 541                 trace_rxrpc_tx_packet(call->debug_id, &txb->wire,
 542                                       rxrpc_tx_point_call_data_frag);
 543         }
 544         rxrpc_tx_backoff(call, ret);
 545
 546         up_write(&conn->local->defrag_sem);
 547         goto done;
 548 }
 549
 550 /*
 551  * Transmit a connection-level abort.
 552  */
 553 void rxrpc_send_conn_abort(struct rxrpc_connection *conn)
 554 {
 555         struct rxrpc_wire_header whdr;
 556         struct msghdr msg;
 557         struct kvec iov[2];
 558         __be32 word;
 559         size_t len;
 560         u32 serial;
 561         int ret;
 562
 563         msg.msg_name    = &conn->peer->srx.transport;
 564         msg.msg_namelen = conn->peer->srx.transport_len;
 565         msg.msg_control = NULL;
 566         msg.msg_controllen = 0;
 567         msg.msg_flags   = 0;
 568
 569         whdr.epoch      = htonl(conn->proto.epoch);
 570         whdr.cid        = htonl(conn->proto.cid);
 571         whdr.callNumber = 0;
 572         whdr.seq        = 0;
 573         whdr.type       = RXRPC_PACKET_TYPE_ABORT;
 574         whdr.flags      = conn->out_clientflag;
 575         whdr.userStatus = 0;
 576         whdr.securityIndex = conn->security_ix;
 577         whdr._rsvd      = 0;
 578         whdr.serviceId  = htons(conn->service_id);
 579
 580         word            = htonl(conn->abort_code);
 581
 582         iov[0].iov_base = &whdr;
 583         iov[0].iov_len  = sizeof(whdr);
 584         iov[1].iov_base = &word;
 585         iov[1].iov_len  = sizeof(word);
 586
 587         len = iov[0].iov_len + iov[1].iov_len;
 588
 589         serial = atomic_inc_return(&conn->serial);
 590         whdr.serial = htonl(serial);
 591
 592         iov_iter_kvec(&msg.msg_iter, WRITE, iov, 2, len);
 593         ret = do_udp_sendmsg(conn->local->socket, &msg, len);
 594         if (ret < 0) {
 595                 trace_rxrpc_tx_fail(conn->debug_id, serial, ret,
 596                                     rxrpc_tx_point_conn_abort);
 597                 _debug("sendmsg failed: %d", ret);
 598                 return;
 599         }
 600
 601         trace_rxrpc_tx_packet(conn->debug_id, &whdr, rxrpc_tx_point_conn_abort);
 602
 603         conn->peer->last_tx_at = ktime_get_seconds();
 604 }
 605
 606 /*
 607  * Reject a packet through the local endpoint.
 608  */
 609 void rxrpc_reject_packet(struct rxrpc_local *local, struct sk_buff *skb)
 610 {
 611         struct rxrpc_wire_header whdr;
 612         struct sockaddr_rxrpc srx;
 613         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 614         struct msghdr msg;
 615         struct kvec iov[2];
 616         size_t size;
 617         __be32 code;
 618         int ret, ioc;
 619
 620         rxrpc_see_skb(skb, rxrpc_skb_see_reject);
 621
 622         iov[0].iov_base = &whdr;
 623         iov[0].iov_len = sizeof(whdr);
 624         iov[1].iov_base = &code;
 625         iov[1].iov_len = sizeof(code);
 626
 627         msg.msg_name = &srx.transport;
 628         msg.msg_control = NULL;
 629         msg.msg_controllen = 0;
 630         msg.msg_flags = 0;
 631
 632         memset(&whdr, 0, sizeof(whdr));
 633
 634         switch (skb->mark) {
 635         case RXRPC_SKB_MARK_REJECT_BUSY:
 636                 whdr.type = RXRPC_PACKET_TYPE_BUSY;
 637                 size = sizeof(whdr);
 638                 ioc = 1;
 639                 break;
 640         case RXRPC_SKB_MARK_REJECT_ABORT:
 641                 whdr.type = RXRPC_PACKET_TYPE_ABORT;
 642                 code = htonl(skb->priority);
 643                 size = sizeof(whdr) + sizeof(code);
 644                 ioc = 2;
 645                 break;
 646         default:
 647                 return;
 648         }
 649
 650         if (rxrpc_extract_addr_from_skb(&srx, skb) == 0) {
 651                 msg.msg_namelen = srx.transport_len;
 652
 653                 whdr.epoch      = htonl(sp->hdr.epoch);
 654                 whdr.cid        = htonl(sp->hdr.cid);
 655                 whdr.callNumber = htonl(sp->hdr.callNumber);
 656                 whdr.serviceId  = htons(sp->hdr.serviceId);
 657                 whdr.flags      = sp->hdr.flags;
 658                 whdr.flags      ^= RXRPC_CLIENT_INITIATED;
 659                 whdr.flags      &= RXRPC_CLIENT_INITIATED;
 660
 661                 iov_iter_kvec(&msg.msg_iter, WRITE, iov, ioc, size);
 662                 ret = do_udp_sendmsg(local->socket, &msg, size);
 663                 if (ret < 0)
 664                         trace_rxrpc_tx_fail(local->debug_id, 0, ret,
 665                                             rxrpc_tx_point_reject);
 666                 else
 667                         trace_rxrpc_tx_packet(local->debug_id, &whdr,
 668                                               rxrpc_tx_point_reject);
 669         }
 670 }
 671
 672 /*
 673  * Send a VERSION reply to a peer as a keepalive.
 674  */
 675 void rxrpc_send_keepalive(struct rxrpc_peer *peer)
 676 {
 677         struct rxrpc_wire_header whdr;
 678         struct msghdr msg;
 679         struct kvec iov[2];
 680         size_t len;
 681         int ret;
 682
 683         _enter("");
 684
 685         msg.msg_name    = &peer->srx.transport;
 686         msg.msg_namelen = peer->srx.transport_len;
 687         msg.msg_control = NULL;
 688         msg.msg_controllen = 0;
 689         msg.msg_flags   = 0;
 690
 691         whdr.epoch      = htonl(peer->local->rxnet->epoch);
 692         whdr.cid        = 0;
 693         whdr.callNumber = 0;
 694         whdr.seq        = 0;
 695         whdr.serial     = 0;
 696         whdr.type       = RXRPC_PACKET_TYPE_VERSION; /* Not client-initiated */
 697         whdr.flags      = RXRPC_LAST_PACKET;
 698         whdr.userStatus = 0;
 699         whdr.securityIndex = 0;
 700         whdr._rsvd      = 0;
 701         whdr.serviceId  = 0;
 702
 703         iov[0].iov_base = &whdr;
 704         iov[0].iov_len  = sizeof(whdr);
 705         iov[1].iov_base = (char *)rxrpc_keepalive_string;
 706         iov[1].iov_len  = sizeof(rxrpc_keepalive_string);
 707
 708         len = iov[0].iov_len + iov[1].iov_len;
 709
 710         iov_iter_kvec(&msg.msg_iter, WRITE, iov, 2, len);
 711         ret = do_udp_sendmsg(peer->local->socket, &msg, len);
 712         if (ret < 0)
 713                 trace_rxrpc_tx_fail(peer->debug_id, 0, ret,
 714                                     rxrpc_tx_point_version_keepalive);
 715         else
 716                 trace_rxrpc_tx_packet(peer->debug_id, &whdr,
 717                                       rxrpc_tx_point_version_keepalive);
 718
 719         peer->last_tx_at = ktime_get_seconds();
 720         _leave("");
 721 }
 722
 723 /*
 724  * Schedule an instant Tx resend.
 725  */
 726 static inline void rxrpc_instant_resend(struct rxrpc_call *call,
 727                                         struct rxrpc_txbuf *txb)
 728 {
 729         if (!__rxrpc_call_is_complete(call))
 730                 kdebug("resend");
 731 }
 732
 733 /*
 734  * Transmit one packet.
 735  */
 736 void rxrpc_transmit_one(struct rxrpc_call *call, struct rxrpc_txbuf *txb)
 737 {
 738         int ret;
 739
 740         ret = rxrpc_send_data_packet(call, txb);
 741         if (ret < 0) {
 742                 switch (ret) {
 743                 case -ENETUNREACH:
 744                 case -EHOSTUNREACH:
 745                 case -ECONNREFUSED:
 746                         rxrpc_set_call_completion(call, RXRPC_CALL_LOCAL_ERROR,
 747                                                   0, ret);
 748                         break;
 749                 default:
 750                         _debug("need instant resend %d", ret);
 751                         rxrpc_instant_resend(call, txb);
 752                 }
 753         } else {
 754                 unsigned long now = jiffies;
 755                 unsigned long resend_at = now + call->peer->rto_j;
 756
 757                 WRITE_ONCE(call->resend_at, resend_at);
 758                 rxrpc_reduce_call_timer(call, resend_at, now,
 759                                         rxrpc_timer_set_for_send);
 760         }
 761 }