net/rxrpc/output.c

   1 /* RxRPC packet transmission
   2  *
   3  * Copyright (C) 2007 Red Hat, Inc. All Rights Reserved.
   4  * Written by David Howells (dhowells@redhat.com)
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  13
  14 #include <linux/net.h>
  15 #include <linux/gfp.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/export.h>
  18 #include <net/sock.h>
  19 #include <net/af_rxrpc.h>
  20 #include "ar-internal.h"
  21
  22 struct rxrpc_ack_buffer {
  23         struct rxrpc_wire_header whdr;
  24         struct rxrpc_ackpacket ack;
  25         u8 acks[255];
  26         u8 pad[3];
  27         struct rxrpc_ackinfo ackinfo;
  28 };
  29
  30 struct rxrpc_abort_buffer {
  31         struct rxrpc_wire_header whdr;
  32         __be32 abort_code;
  33 };
  34
  35 static const char rxrpc_keepalive_string[] = "";
  36
  37 /*
  38  * Arrange for a keepalive ping a certain time after we last transmitted.  This
  39  * lets the far side know we're still interested in this call and helps keep
  40  * the route through any intervening firewall open.
  41  *
  42  * Receiving a response to the ping will prevent the ->expect_rx_by timer from
  43  * expiring.
  44  */
  45 static void rxrpc_set_keepalive(struct rxrpc_call *call)
  46 {
  47         unsigned long now = jiffies, keepalive_at = call->next_rx_timo / 6;
  48
  49         keepalive_at += now;
  50         WRITE_ONCE(call->keepalive_at, keepalive_at);
  51         rxrpc_reduce_call_timer(call, keepalive_at, now,
  52                                 rxrpc_timer_set_for_keepalive);
  53 }
  54
  55 /*
  56  * Fill out an ACK packet.
  57  */
  58 static size_t rxrpc_fill_out_ack(struct rxrpc_connection *conn,
  59                                  struct rxrpc_call *call,
  60                                  struct rxrpc_ack_buffer *pkt,
  61                                  rxrpc_seq_t *_hard_ack,
  62                                  rxrpc_seq_t *_top,
  63                                  u8 reason)
  64 {
  65         rxrpc_serial_t serial;
  66         rxrpc_seq_t hard_ack, top, seq;
  67         int ix;
  68         u32 mtu, jmax;
  69         u8 *ackp = pkt->acks;
  70
  71         /* Barrier against rxrpc_input_data(). */
  72         serial = call->ackr_serial;
  73         hard_ack = READ_ONCE(call->rx_hard_ack);
  74         top = smp_load_acquire(&call->rx_top);
  75         *_hard_ack = hard_ack;
  76         *_top = top;
  77
  78         pkt->ack.bufferSpace    = htons(8);
  79         pkt->ack.maxSkew        = htons(call->ackr_skew);
  80         pkt->ack.firstPacket    = htonl(hard_ack + 1);
  81         pkt->ack.previousPacket = htonl(call->ackr_prev_seq);
  82         pkt->ack.serial         = htonl(serial);
  83         pkt->ack.reason         = reason;
  84         pkt->ack.nAcks          = top - hard_ack;
  85
  86         if (reason == RXRPC_ACK_PING)
  87                 pkt->whdr.flags |= RXRPC_REQUEST_ACK;
  88
  89         if (after(top, hard_ack)) {
  90                 seq = hard_ack + 1;
  91                 do {
  92                         ix = seq & RXRPC_RXTX_BUFF_MASK;
  93                         if (call->rxtx_buffer[ix])
  94                                 *ackp++ = RXRPC_ACK_TYPE_ACK;
  95                         else
  96                                 *ackp++ = RXRPC_ACK_TYPE_NACK;
  97                         seq++;
  98                 } while (before_eq(seq, top));
  99         }
 100
 101         mtu = conn->params.peer->if_mtu;
 102         mtu -= conn->params.peer->hdrsize;
 103         jmax = (call->nr_jumbo_bad > 3) ? 1 : rxrpc_rx_jumbo_max;
 104         pkt->ackinfo.rxMTU      = htonl(rxrpc_rx_mtu);
 105         pkt->ackinfo.maxMTU     = htonl(mtu);
 106         pkt->ackinfo.rwind      = htonl(call->rx_winsize);
 107         pkt->ackinfo.jumbo_max  = htonl(jmax);
 108
 109         *ackp++ = 0;
 110         *ackp++ = 0;
 111         *ackp++ = 0;
 112         return top - hard_ack + 3;
 113 }
 114
 115 /*
 116  * Send an ACK call packet.
 117  */
 118 int rxrpc_send_ack_packet(struct rxrpc_call *call, bool ping,
 119                           rxrpc_serial_t *_serial)
 120 {
 121         struct rxrpc_connection *conn = NULL;
 122         struct rxrpc_ack_buffer *pkt;
 123         struct msghdr msg;
 124         struct kvec iov[2];
 125         rxrpc_serial_t serial;
 126         rxrpc_seq_t hard_ack, top;
 127         ktime_t now;
 128         size_t len, n;
 129         int ret;
 130         u8 reason;
 131
 132         spin_lock_bh(&call->lock);
 133         if (call->conn)
 134                 conn = rxrpc_get_connection_maybe(call->conn);
 135         spin_unlock_bh(&call->lock);
 136         if (!conn)
 137                 return -ECONNRESET;
 138
 139         pkt = kzalloc(sizeof(*pkt), GFP_KERNEL);
 140         if (!pkt) {
 141                 rxrpc_put_connection(conn);
 142                 return -ENOMEM;
 143         }
 144
 145         msg.msg_name    = &call->peer->srx.transport;
 146         msg.msg_namelen = call->peer->srx.transport_len;
 147         msg.msg_control = NULL;
 148         msg.msg_controllen = 0;
 149         msg.msg_flags   = 0;
 150
 151         pkt->whdr.epoch         = htonl(conn->proto.epoch);
 152         pkt->whdr.cid           = htonl(call->cid);
 153         pkt->whdr.callNumber    = htonl(call->call_id);
 154         pkt->whdr.seq           = 0;
 155         pkt->whdr.type          = RXRPC_PACKET_TYPE_ACK;
 156         pkt->whdr.flags         = RXRPC_SLOW_START_OK | conn->out_clientflag;
 157         pkt->whdr.userStatus    = 0;
 158         pkt->whdr.securityIndex = call->security_ix;
 159         pkt->whdr._rsvd         = 0;
 160         pkt->whdr.serviceId     = htons(call->service_id);
 161
 162         spin_lock_bh(&call->lock);
 163         if (ping) {
 164                 reason = RXRPC_ACK_PING;
 165         } else {
 166                 reason = call->ackr_reason;
 167                 if (!call->ackr_reason) {
 168                         spin_unlock_bh(&call->lock);
 169                         ret = 0;
 170                         goto out;
 171                 }
 172                 call->ackr_reason = 0;
 173         }
 174         n = rxrpc_fill_out_ack(conn, call, pkt, &hard_ack, &top, reason);
 175
 176         spin_unlock_bh(&call->lock);
 177
 178         iov[0].iov_base = pkt;
 179         iov[0].iov_len  = sizeof(pkt->whdr) + sizeof(pkt->ack) + n;
 180         iov[1].iov_base = &pkt->ackinfo;
 181         iov[1].iov_len  = sizeof(pkt->ackinfo);
 182         len = iov[0].iov_len + iov[1].iov_len;
 183
 184         serial = atomic_inc_return(&conn->serial);
 185         pkt->whdr.serial = htonl(serial);
 186         trace_rxrpc_tx_ack(call, serial,
 187                            ntohl(pkt->ack.firstPacket),
 188                            ntohl(pkt->ack.serial),
 189                            pkt->ack.reason, pkt->ack.nAcks);
 190         if (_serial)
 191                 *_serial = serial;
 192
 193         if (ping) {
 194                 call->ping_serial = serial;
 195                 smp_wmb();
 196                 /* We need to stick a time in before we send the packet in case
 197                  * the reply gets back before kernel_sendmsg() completes - but
 198                  * asking UDP to send the packet can take a relatively long
 199                  * time, so we update the time after, on the assumption that
 200                  * the packet transmission is more likely to happen towards the
 201                  * end of the kernel_sendmsg() call.
 202                  */
 203                 call->ping_time = ktime_get_real();
 204                 set_bit(RXRPC_CALL_PINGING, &call->flags);
 205                 trace_rxrpc_rtt_tx(call, rxrpc_rtt_tx_ping, serial);
 206         }
 207
 208         ret = kernel_sendmsg(conn->params.local->socket, &msg, iov, 2, len);
 209         now = ktime_get_real();
 210         if (ping)
 211                 call->ping_time = now;
 212         conn->params.peer->last_tx_at = ktime_get_real();
 213
 214         if (call->state < RXRPC_CALL_COMPLETE) {
 215                 if (ret < 0) {
 216                         if (ping)
 217                                 clear_bit(RXRPC_CALL_PINGING, &call->flags);
 218                         rxrpc_propose_ACK(call, pkt->ack.reason,
 219                                           ntohs(pkt->ack.maxSkew),
 220                                           ntohl(pkt->ack.serial),
 221                                           true, true,
 222                                           rxrpc_propose_ack_retry_tx);
 223                 } else {
 224                         spin_lock_bh(&call->lock);
 225                         if (after(hard_ack, call->ackr_consumed))
 226                                 call->ackr_consumed = hard_ack;
 227                         if (after(top, call->ackr_seen))
 228                                 call->ackr_seen = top;
 229                         spin_unlock_bh(&call->lock);
 230                 }
 231
 232                 rxrpc_set_keepalive(call);
 233         }
 234
 235 out:
 236         rxrpc_put_connection(conn);
 237         kfree(pkt);
 238         return ret;
 239 }
 240
 241 /*
 242  * Send an ABORT call packet.
 243  */
 244 int rxrpc_send_abort_packet(struct rxrpc_call *call)
 245 {
 246         struct rxrpc_connection *conn = NULL;
 247         struct rxrpc_abort_buffer pkt;
 248         struct msghdr msg;
 249         struct kvec iov[1];
 250         rxrpc_serial_t serial;
 251         int ret;
 252
 253         /* Don't bother sending aborts for a client call once the server has
 254          * hard-ACK'd all of its request data.  After that point, we're not
 255          * going to stop the operation proceeding, and whilst we might limit
 256          * the reply, it's not worth it if we can send a new call on the same
 257          * channel instead, thereby closing off this call.
 258          */
 259         if (rxrpc_is_client_call(call) &&
 260             test_bit(RXRPC_CALL_TX_LAST, &call->flags))
 261                 return 0;
 262
 263         spin_lock_bh(&call->lock);
 264         if (call->conn)
 265                 conn = rxrpc_get_connection_maybe(call->conn);
 266         spin_unlock_bh(&call->lock);
 267         if (!conn)
 268                 return -ECONNRESET;
 269
 270         msg.msg_name    = &call->peer->srx.transport;
 271         msg.msg_namelen = call->peer->srx.transport_len;
 272         msg.msg_control = NULL;
 273         msg.msg_controllen = 0;
 274         msg.msg_flags   = 0;
 275
 276         pkt.whdr.epoch          = htonl(conn->proto.epoch);
 277         pkt.whdr.cid            = htonl(call->cid);
 278         pkt.whdr.callNumber     = htonl(call->call_id);
 279         pkt.whdr.seq            = 0;
 280         pkt.whdr.type           = RXRPC_PACKET_TYPE_ABORT;
 281         pkt.whdr.flags          = conn->out_clientflag;
 282         pkt.whdr.userStatus     = 0;
 283         pkt.whdr.securityIndex  = call->security_ix;
 284         pkt.whdr._rsvd          = 0;
 285         pkt.whdr.serviceId      = htons(call->service_id);
 286         pkt.abort_code          = htonl(call->abort_code);
 287
 288         iov[0].iov_base = &pkt;
 289         iov[0].iov_len  = sizeof(pkt);
 290
 291         serial = atomic_inc_return(&conn->serial);
 292         pkt.whdr.serial = htonl(serial);
 293
 294         ret = kernel_sendmsg(conn->params.local->socket,
 295                              &msg, iov, 1, sizeof(pkt));
 296         conn->params.peer->last_tx_at = ktime_get_real();
 297
 298         rxrpc_put_connection(conn);
 299         return ret;
 300 }
 301
 302 /*
 303  * send a packet through the transport endpoint
 304  */
 305 int rxrpc_send_data_packet(struct rxrpc_call *call, struct sk_buff *skb,
 306                            bool retrans)
 307 {
 308         struct rxrpc_connection *conn = call->conn;
 309         struct rxrpc_wire_header whdr;
 310         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 311         struct msghdr msg;
 312         struct kvec iov[2];
 313         rxrpc_serial_t serial;
 314         size_t len;
 315         bool lost = false;
 316         int ret, opt;
 317
 318         _enter(",{%d}", skb->len);
 319
 320         /* Each transmission of a Tx packet needs a new serial number */
 321         serial = atomic_inc_return(&conn->serial);
 322
 323         whdr.epoch      = htonl(conn->proto.epoch);
 324         whdr.cid        = htonl(call->cid);
 325         whdr.callNumber = htonl(call->call_id);
 326         whdr.seq        = htonl(sp->hdr.seq);
 327         whdr.serial     = htonl(serial);
 328         whdr.type       = RXRPC_PACKET_TYPE_DATA;
 329         whdr.flags      = sp->hdr.flags;
 330         whdr.userStatus = 0;
 331         whdr.securityIndex = call->security_ix;
 332         whdr._rsvd      = htons(sp->hdr._rsvd);
 333         whdr.serviceId  = htons(call->service_id);
 334
 335         if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags) &&
 336             sp->hdr.seq == 1)
 337                 whdr.userStatus = RXRPC_USERSTATUS_SERVICE_UPGRADE;
 338
 339         iov[0].iov_base = &whdr;
 340         iov[0].iov_len = sizeof(whdr);
 341         iov[1].iov_base = skb->head;
 342         iov[1].iov_len = skb->len;
 343         len = iov[0].iov_len + iov[1].iov_len;
 344
 345         msg.msg_name = &call->peer->srx.transport;
 346         msg.msg_namelen = call->peer->srx.transport_len;
 347         msg.msg_control = NULL;
 348         msg.msg_controllen = 0;
 349         msg.msg_flags = 0;
 350
 351         /* If our RTT cache needs working on, request an ACK.  Also request
 352          * ACKs if a DATA packet appears to have been lost.
 353          */
 354         if (!(sp->hdr.flags & RXRPC_LAST_PACKET) &&
 355             (test_and_clear_bit(RXRPC_CALL_EV_ACK_LOST, &call->events) ||
 356              retrans ||
 357              call->cong_mode == RXRPC_CALL_SLOW_START ||
 358              (call->peer->rtt_usage < 3 && sp->hdr.seq & 1) ||
 359              ktime_before(ktime_add_ms(call->peer->rtt_last_req, 1000),
 360                           ktime_get_real())))
 361                 whdr.flags |= RXRPC_REQUEST_ACK;
 362
 363         if (IS_ENABLED(CONFIG_AF_RXRPC_INJECT_LOSS)) {
 364                 static int lose;
 365                 if ((lose++ & 7) == 7) {
 366                         ret = 0;
 367                         lost = true;
 368                         goto done;
 369                 }
 370         }
 371
 372         _proto("Tx DATA %%%u { #%u }", serial, sp->hdr.seq);
 373
 374         /* send the packet with the don't fragment bit set if we currently
 375          * think it's small enough */
 376         if (iov[1].iov_len >= call->peer->maxdata)
 377                 goto send_fragmentable;
 378
 379         down_read(&conn->params.local->defrag_sem);
 380         /* send the packet by UDP
 381          * - returns -EMSGSIZE if UDP would have to fragment the packet
 382          *   to go out of the interface
 383          *   - in which case, we'll have processed the ICMP error
 384          *     message and update the peer record
 385          */
 386         ret = kernel_sendmsg(conn->params.local->socket, &msg, iov, 2, len);
 387         conn->params.peer->last_tx_at = ktime_get_real();
 388
 389         up_read(&conn->params.local->defrag_sem);
 390         if (ret == -EMSGSIZE)
 391                 goto send_fragmentable;
 392
 393 done:
 394         trace_rxrpc_tx_data(call, sp->hdr.seq, serial, whdr.flags,
 395                             retrans, lost);
 396         if (ret >= 0) {
 397                 ktime_t now = ktime_get_real();
 398                 skb->tstamp = now;
 399                 smp_wmb();
 400                 sp->hdr.serial = serial;
 401                 if (whdr.flags & RXRPC_REQUEST_ACK) {
 402                         call->peer->rtt_last_req = now;
 403                         trace_rxrpc_rtt_tx(call, rxrpc_rtt_tx_data, serial);
 404                         if (call->peer->rtt_usage > 1) {
 405                                 unsigned long nowj = jiffies, ack_lost_at;
 406
 407                                 ack_lost_at = nsecs_to_jiffies(2 * call->peer->rtt);
 408                                 if (ack_lost_at < 1)
 409                                         ack_lost_at = 1;
 410
 411                                 ack_lost_at += nowj;
 412                                 WRITE_ONCE(call->ack_lost_at, ack_lost_at);
 413                                 rxrpc_reduce_call_timer(call, ack_lost_at, nowj,
 414                                                         rxrpc_timer_set_for_lost_ack);
 415                         }
 416                 }
 417         }
 418
 419         rxrpc_set_keepalive(call);
 420
 421         _leave(" = %d [%u]", ret, call->peer->maxdata);
 422         return ret;
 423
 424 send_fragmentable:
 425         /* attempt to send this message with fragmentation enabled */
 426         _debug("send fragment");
 427
 428         down_write(&conn->params.local->defrag_sem);
 429
 430         switch (conn->params.local->srx.transport.family) {
 431         case AF_INET:
 432                 opt = IP_PMTUDISC_DONT;
 433                 ret = kernel_setsockopt(conn->params.local->socket,
 434                                         SOL_IP, IP_MTU_DISCOVER,
 435                                         (char *)&opt, sizeof(opt));
 436                 if (ret == 0) {
 437                         ret = kernel_sendmsg(conn->params.local->socket, &msg,
 438                                              iov, 2, len);
 439                         conn->params.peer->last_tx_at = ktime_get_real();
 440
 441                         opt = IP_PMTUDISC_DO;
 442                         kernel_setsockopt(conn->params.local->socket, SOL_IP,
 443                                           IP_MTU_DISCOVER,
 444                                           (char *)&opt, sizeof(opt));
 445                 }
 446                 break;
 447
 448 #ifdef CONFIG_AF_RXRPC_IPV6
 449         case AF_INET6:
 450                 opt = IPV6_PMTUDISC_DONT;
 451                 ret = kernel_setsockopt(conn->params.local->socket,
 452                                         SOL_IPV6, IPV6_MTU_DISCOVER,
 453                                         (char *)&opt, sizeof(opt));
 454                 if (ret == 0) {
 455                         ret = kernel_sendmsg(conn->params.local->socket, &msg,
 456                                              iov, 2, len);
 457                         conn->params.peer->last_tx_at = ktime_get_real();
 458
 459                         opt = IPV6_PMTUDISC_DO;
 460                         kernel_setsockopt(conn->params.local->socket,
 461                                           SOL_IPV6, IPV6_MTU_DISCOVER,
 462                                           (char *)&opt, sizeof(opt));
 463                 }
 464                 break;
 465 #endif
 466         }
 467
 468         up_write(&conn->params.local->defrag_sem);
 469         goto done;
 470 }
 471
 472 /*
 473  * reject packets through the local endpoint
 474  */
 475 void rxrpc_reject_packets(struct rxrpc_local *local)
 476 {
 477         struct sockaddr_rxrpc srx;
 478         struct rxrpc_skb_priv *sp;
 479         struct rxrpc_wire_header whdr;
 480         struct sk_buff *skb;
 481         struct msghdr msg;
 482         struct kvec iov[2];
 483         size_t size;
 484         __be32 code;
 485
 486         _enter("%d", local->debug_id);
 487
 488         iov[0].iov_base = &whdr;
 489         iov[0].iov_len = sizeof(whdr);
 490         iov[1].iov_base = &code;
 491         iov[1].iov_len = sizeof(code);
 492         size = sizeof(whdr) + sizeof(code);
 493
 494         msg.msg_name = &srx.transport;
 495         msg.msg_control = NULL;
 496         msg.msg_controllen = 0;
 497         msg.msg_flags = 0;
 498
 499         memset(&whdr, 0, sizeof(whdr));
 500         whdr.type = RXRPC_PACKET_TYPE_ABORT;
 501
 502         while ((skb = skb_dequeue(&local->reject_queue))) {
 503                 rxrpc_see_skb(skb, rxrpc_skb_rx_seen);
 504                 sp = rxrpc_skb(skb);
 505
 506                 if (rxrpc_extract_addr_from_skb(local, &srx, skb) == 0) {
 507                         msg.msg_namelen = srx.transport_len;
 508
 509                         code = htonl(skb->priority);
 510
 511                         whdr.epoch      = htonl(sp->hdr.epoch);
 512                         whdr.cid        = htonl(sp->hdr.cid);
 513                         whdr.callNumber = htonl(sp->hdr.callNumber);
 514                         whdr.serviceId  = htons(sp->hdr.serviceId);
 515                         whdr.flags      = sp->hdr.flags;
 516                         whdr.flags      ^= RXRPC_CLIENT_INITIATED;
 517                         whdr.flags      &= RXRPC_CLIENT_INITIATED;
 518
 519                         kernel_sendmsg(local->socket, &msg, iov, 2, size);
 520                 }
 521
 522                 rxrpc_free_skb(skb, rxrpc_skb_rx_freed);
 523         }
 524
 525         _leave("");
 526 }
 527
 528 /*
 529  * Send a VERSION reply to a peer as a keepalive.
 530  */
 531 void rxrpc_send_keepalive(struct rxrpc_peer *peer)
 532 {
 533         struct rxrpc_wire_header whdr;
 534         struct msghdr msg;
 535         struct kvec iov[2];
 536         size_t len;
 537         int ret;
 538
 539         _enter("");
 540
 541         msg.msg_name    = &peer->srx.transport;
 542         msg.msg_namelen = peer->srx.transport_len;
 543         msg.msg_control = NULL;
 544         msg.msg_controllen = 0;
 545         msg.msg_flags   = 0;
 546
 547         whdr.epoch      = htonl(peer->local->rxnet->epoch);
 548         whdr.cid        = 0;
 549         whdr.callNumber = 0;
 550         whdr.seq        = 0;
 551         whdr.serial     = 0;
 552         whdr.type       = RXRPC_PACKET_TYPE_VERSION; /* Not client-initiated */
 553         whdr.flags      = RXRPC_LAST_PACKET;
 554         whdr.userStatus = 0;
 555         whdr.securityIndex = 0;
 556         whdr._rsvd      = 0;
 557         whdr.serviceId  = 0;
 558
 559         iov[0].iov_base = &whdr;
 560         iov[0].iov_len  = sizeof(whdr);
 561         iov[1].iov_base = (char *)rxrpc_keepalive_string;
 562         iov[1].iov_len  = sizeof(rxrpc_keepalive_string);
 563
 564         len = iov[0].iov_len + iov[1].iov_len;
 565
 566         _proto("Tx VERSION (keepalive)");
 567
 568         ret = kernel_sendmsg(peer->local->socket, &msg, iov, 2, len);
 569         if (ret < 0)
 570                 _debug("sendmsg failed: %d", ret);
 571
 572         peer->last_tx_at = ktime_get_real();
 573         _leave("");
 574 }