net/rxrpc/input.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* Processing of received RxRPC packets
   3  *
   4  * Copyright (C) 2020 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 "ar-internal.h"
  11
  12 static void rxrpc_proto_abort(const char *why,
  13                               struct rxrpc_call *call, rxrpc_seq_t seq)
  14 {
  15         if (rxrpc_abort_call(why, call, seq, RX_PROTOCOL_ERROR, -EBADMSG))
  16                 rxrpc_send_abort_packet(call);
  17 }
  18
  19 /*
  20  * Do TCP-style congestion management [RFC 5681].
  21  */
  22 static void rxrpc_congestion_management(struct rxrpc_call *call,
  23                                         struct sk_buff *skb,
  24                                         struct rxrpc_ack_summary *summary,
  25                                         rxrpc_serial_t acked_serial)
  26 {
  27         enum rxrpc_congest_change change = rxrpc_cong_no_change;
  28         unsigned int cumulative_acks = call->cong_cumul_acks;
  29         unsigned int cwnd = call->cong_cwnd;
  30         bool resend = false;
  31
  32         summary->flight_size =
  33                 (call->tx_top - call->acks_hard_ack) - summary->nr_acks;
  34
  35         if (test_and_clear_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags)) {
  36                 summary->retrans_timeo = true;
  37                 call->cong_ssthresh = max_t(unsigned int,
  38                                             summary->flight_size / 2, 2);
  39                 cwnd = 1;
  40                 if (cwnd >= call->cong_ssthresh &&
  41                     call->cong_mode == RXRPC_CALL_SLOW_START) {
  42                         call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
  43                         call->cong_tstamp = skb->tstamp;
  44                         cumulative_acks = 0;
  45                 }
  46         }
  47
  48         cumulative_acks += summary->nr_new_acks;
  49         cumulative_acks += summary->nr_rot_new_acks;
  50         if (cumulative_acks > 255)
  51                 cumulative_acks = 255;
  52
  53         summary->mode = call->cong_mode;
  54         summary->cwnd = call->cong_cwnd;
  55         summary->ssthresh = call->cong_ssthresh;
  56         summary->cumulative_acks = cumulative_acks;
  57         summary->dup_acks = call->cong_dup_acks;
  58
  59         /* If we haven't transmitted anything for >1RTT, we should reset the
  60          * congestion management state.
  61          */
  62         if ((call->cong_mode == RXRPC_CALL_SLOW_START ||
  63              call->cong_mode == RXRPC_CALL_CONGEST_AVOIDANCE) &&
  64             ktime_before(ktime_add_us(call->tx_last_sent,
  65                                       call->peer->srtt_us >> 3),
  66                          ktime_get_real())
  67             ) {
  68                 change = rxrpc_cong_idle_reset;
  69                 summary->mode = RXRPC_CALL_SLOW_START;
  70                 if (RXRPC_TX_SMSS > 2190)
  71                         summary->cwnd = 2;
  72                 else if (RXRPC_TX_SMSS > 1095)
  73                         summary->cwnd = 3;
  74                 else
  75                         summary->cwnd = 4;
  76         }
  77
  78         switch (call->cong_mode) {
  79         case RXRPC_CALL_SLOW_START:
  80                 if (summary->saw_nacks)
  81                         goto packet_loss_detected;
  82                 if (summary->cumulative_acks > 0)
  83                         cwnd += 1;
  84                 if (cwnd >= call->cong_ssthresh) {
  85                         call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
  86                         call->cong_tstamp = skb->tstamp;
  87                 }
  88                 goto out;
  89
  90         case RXRPC_CALL_CONGEST_AVOIDANCE:
  91                 if (summary->saw_nacks)
  92                         goto packet_loss_detected;
  93
  94                 /* We analyse the number of packets that get ACK'd per RTT
  95                  * period and increase the window if we managed to fill it.
  96                  */
  97                 if (call->peer->rtt_count == 0)
  98                         goto out;
  99                 if (ktime_before(skb->tstamp,
 100                                  ktime_add_us(call->cong_tstamp,
 101                                               call->peer->srtt_us >> 3)))
 102                         goto out_no_clear_ca;
 103                 change = rxrpc_cong_rtt_window_end;
 104                 call->cong_tstamp = skb->tstamp;
 105                 if (cumulative_acks >= cwnd)
 106                         cwnd++;
 107                 goto out;
 108
 109         case RXRPC_CALL_PACKET_LOSS:
 110                 if (!summary->saw_nacks)
 111                         goto resume_normality;
 112
 113                 if (summary->new_low_nack) {
 114                         change = rxrpc_cong_new_low_nack;
 115                         call->cong_dup_acks = 1;
 116                         if (call->cong_extra > 1)
 117                                 call->cong_extra = 1;
 118                         goto send_extra_data;
 119                 }
 120
 121                 call->cong_dup_acks++;
 122                 if (call->cong_dup_acks < 3)
 123                         goto send_extra_data;
 124
 125                 change = rxrpc_cong_begin_retransmission;
 126                 call->cong_mode = RXRPC_CALL_FAST_RETRANSMIT;
 127                 call->cong_ssthresh = max_t(unsigned int,
 128                                             summary->flight_size / 2, 2);
 129                 cwnd = call->cong_ssthresh + 3;
 130                 call->cong_extra = 0;
 131                 call->cong_dup_acks = 0;
 132                 resend = true;
 133                 goto out;
 134
 135         case RXRPC_CALL_FAST_RETRANSMIT:
 136                 if (!summary->new_low_nack) {
 137                         if (summary->nr_new_acks == 0)
 138                                 cwnd += 1;
 139                         call->cong_dup_acks++;
 140                         if (call->cong_dup_acks == 2) {
 141                                 change = rxrpc_cong_retransmit_again;
 142                                 call->cong_dup_acks = 0;
 143                                 resend = true;
 144                         }
 145                 } else {
 146                         change = rxrpc_cong_progress;
 147                         cwnd = call->cong_ssthresh;
 148                         if (!summary->saw_nacks)
 149                                 goto resume_normality;
 150                 }
 151                 goto out;
 152
 153         default:
 154                 BUG();
 155                 goto out;
 156         }
 157
 158 resume_normality:
 159         change = rxrpc_cong_cleared_nacks;
 160         call->cong_dup_acks = 0;
 161         call->cong_extra = 0;
 162         call->cong_tstamp = skb->tstamp;
 163         if (cwnd < call->cong_ssthresh)
 164                 call->cong_mode = RXRPC_CALL_SLOW_START;
 165         else
 166                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
 167 out:
 168         cumulative_acks = 0;
 169 out_no_clear_ca:
 170         if (cwnd >= RXRPC_TX_MAX_WINDOW)
 171                 cwnd = RXRPC_TX_MAX_WINDOW;
 172         call->cong_cwnd = cwnd;
 173         call->cong_cumul_acks = cumulative_acks;
 174         trace_rxrpc_congest(call, summary, acked_serial, change);
 175         if (resend)
 176                 rxrpc_resend(call, skb);
 177         return;
 178
 179 packet_loss_detected:
 180         change = rxrpc_cong_saw_nack;
 181         call->cong_mode = RXRPC_CALL_PACKET_LOSS;
 182         call->cong_dup_acks = 0;
 183         goto send_extra_data;
 184
 185 send_extra_data:
 186         /* Send some previously unsent DATA if we have some to advance the ACK
 187          * state.
 188          */
 189         if (test_bit(RXRPC_CALL_TX_LAST, &call->flags) ||
 190             summary->nr_acks != call->tx_top - call->acks_hard_ack) {
 191                 call->cong_extra++;
 192                 wake_up(&call->waitq);
 193         }
 194         goto out_no_clear_ca;
 195 }
 196
 197 /*
 198  * Apply a hard ACK by advancing the Tx window.
 199  */
 200 static bool rxrpc_rotate_tx_window(struct rxrpc_call *call, rxrpc_seq_t to,
 201                                    struct rxrpc_ack_summary *summary)
 202 {
 203         struct rxrpc_txbuf *txb;
 204         bool rot_last = false;
 205
 206         list_for_each_entry_rcu(txb, &call->tx_buffer, call_link, false) {
 207                 if (before_eq(txb->seq, call->acks_hard_ack))
 208                         continue;
 209                 summary->nr_rot_new_acks++;
 210                 if (test_bit(RXRPC_TXBUF_LAST, &txb->flags)) {
 211                         set_bit(RXRPC_CALL_TX_LAST, &call->flags);
 212                         rot_last = true;
 213                 }
 214                 if (txb->seq == to)
 215                         break;
 216         }
 217
 218         if (rot_last)
 219                 set_bit(RXRPC_CALL_TX_ALL_ACKED, &call->flags);
 220
 221         _enter("%x,%x,%x,%d", to, call->acks_hard_ack, call->tx_top, rot_last);
 222
 223         if (call->acks_lowest_nak == call->acks_hard_ack) {
 224                 call->acks_lowest_nak = to;
 225         } else if (after(to, call->acks_lowest_nak)) {
 226                 summary->new_low_nack = true;
 227                 call->acks_lowest_nak = to;
 228         }
 229
 230         smp_store_release(&call->acks_hard_ack, to);
 231
 232         trace_rxrpc_txqueue(call, (rot_last ?
 233                                    rxrpc_txqueue_rotate_last :
 234                                    rxrpc_txqueue_rotate));
 235         wake_up(&call->waitq);
 236         return rot_last;
 237 }
 238
 239 /*
 240  * End the transmission phase of a call.
 241  *
 242  * This occurs when we get an ACKALL packet, the first DATA packet of a reply,
 243  * or a final ACK packet.
 244  */
 245 static bool rxrpc_end_tx_phase(struct rxrpc_call *call, bool reply_begun,
 246                                const char *abort_why)
 247 {
 248         unsigned int state;
 249
 250         ASSERT(test_bit(RXRPC_CALL_TX_LAST, &call->flags));
 251
 252         write_lock(&call->state_lock);
 253
 254         state = call->state;
 255         switch (state) {
 256         case RXRPC_CALL_CLIENT_SEND_REQUEST:
 257         case RXRPC_CALL_CLIENT_AWAIT_REPLY:
 258                 if (reply_begun)
 259                         call->state = state = RXRPC_CALL_CLIENT_RECV_REPLY;
 260                 else
 261                         call->state = state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
 262                 break;
 263
 264         case RXRPC_CALL_SERVER_AWAIT_ACK:
 265                 __rxrpc_call_completed(call);
 266                 state = call->state;
 267                 break;
 268
 269         default:
 270                 goto bad_state;
 271         }
 272
 273         write_unlock(&call->state_lock);
 274         if (state == RXRPC_CALL_CLIENT_AWAIT_REPLY)
 275                 trace_rxrpc_txqueue(call, rxrpc_txqueue_await_reply);
 276         else
 277                 trace_rxrpc_txqueue(call, rxrpc_txqueue_end);
 278         _leave(" = ok");
 279         return true;
 280
 281 bad_state:
 282         write_unlock(&call->state_lock);
 283         kdebug("end_tx %s", rxrpc_call_states[call->state]);
 284         rxrpc_proto_abort(abort_why, call, call->tx_top);
 285         return false;
 286 }
 287
 288 /*
 289  * Begin the reply reception phase of a call.
 290  */
 291 static bool rxrpc_receiving_reply(struct rxrpc_call *call)
 292 {
 293         struct rxrpc_ack_summary summary = { 0 };
 294         unsigned long now, timo;
 295         rxrpc_seq_t top = READ_ONCE(call->tx_top);
 296
 297         if (call->ackr_reason) {
 298                 now = jiffies;
 299                 timo = now + MAX_JIFFY_OFFSET;
 300                 WRITE_ONCE(call->resend_at, timo);
 301                 WRITE_ONCE(call->delay_ack_at, timo);
 302                 trace_rxrpc_timer(call, rxrpc_timer_init_for_reply, now);
 303         }
 304
 305         if (!test_bit(RXRPC_CALL_TX_LAST, &call->flags)) {
 306                 if (!rxrpc_rotate_tx_window(call, top, &summary)) {
 307                         rxrpc_proto_abort("TXL", call, top);
 308                         return false;
 309                 }
 310         }
 311         return rxrpc_end_tx_phase(call, true, "ETD");
 312 }
 313
 314 static void rxrpc_input_update_ack_window(struct rxrpc_call *call,
 315                                           rxrpc_seq_t window, rxrpc_seq_t wtop)
 316 {
 317         atomic64_set_release(&call->ackr_window, ((u64)wtop) << 32 | window);
 318 }
 319
 320 /*
 321  * Push a DATA packet onto the Rx queue.
 322  */
 323 static void rxrpc_input_queue_data(struct rxrpc_call *call, struct sk_buff *skb,
 324                                    rxrpc_seq_t window, rxrpc_seq_t wtop,
 325                                    enum rxrpc_receive_trace why)
 326 {
 327         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 328         bool last = sp->hdr.flags & RXRPC_LAST_PACKET;
 329
 330         __skb_queue_tail(&call->recvmsg_queue, skb);
 331         rxrpc_input_update_ack_window(call, window, wtop);
 332
 333         trace_rxrpc_receive(call, last ? why + 1 : why, sp->hdr.serial, sp->hdr.seq);
 334 }
 335
 336 /*
 337  * Process a DATA packet.
 338  */
 339 static void rxrpc_input_data_one(struct rxrpc_call *call, struct sk_buff *skb,
 340                                  bool *_notify)
 341 {
 342         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 343         struct sk_buff *oos;
 344         rxrpc_serial_t serial = sp->hdr.serial;
 345         u64 win = atomic64_read(&call->ackr_window);
 346         rxrpc_seq_t window = lower_32_bits(win);
 347         rxrpc_seq_t wtop = upper_32_bits(win);
 348         rxrpc_seq_t wlimit = window + call->rx_winsize - 1;
 349         rxrpc_seq_t seq = sp->hdr.seq;
 350         bool last = sp->hdr.flags & RXRPC_LAST_PACKET;
 351         int ack_reason = -1;
 352
 353         rxrpc_inc_stat(call->rxnet, stat_rx_data);
 354         if (sp->hdr.flags & RXRPC_REQUEST_ACK)
 355                 rxrpc_inc_stat(call->rxnet, stat_rx_data_reqack);
 356         if (sp->hdr.flags & RXRPC_JUMBO_PACKET)
 357                 rxrpc_inc_stat(call->rxnet, stat_rx_data_jumbo);
 358
 359         if (last) {
 360                 if (test_and_set_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
 361                     seq + 1 != wtop) {
 362                         rxrpc_proto_abort("LSN", call, seq);
 363                         return;
 364                 }
 365         } else {
 366                 if (test_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
 367                     after_eq(seq, wtop)) {
 368                         pr_warn("Packet beyond last: c=%x q=%x window=%x-%x wlimit=%x\n",
 369                                 call->debug_id, seq, window, wtop, wlimit);
 370                         rxrpc_proto_abort("LSA", call, seq);
 371                         return;
 372                 }
 373         }
 374
 375         if (after(seq, call->rx_highest_seq))
 376                 call->rx_highest_seq = seq;
 377
 378         trace_rxrpc_rx_data(call->debug_id, seq, serial, sp->hdr.flags);
 379
 380         if (before(seq, window)) {
 381                 ack_reason = RXRPC_ACK_DUPLICATE;
 382                 goto send_ack;
 383         }
 384         if (after(seq, wlimit)) {
 385                 ack_reason = RXRPC_ACK_EXCEEDS_WINDOW;
 386                 goto send_ack;
 387         }
 388
 389         /* Queue the packet. */
 390         if (seq == window) {
 391                 rxrpc_seq_t reset_from;
 392                 bool reset_sack = false;
 393
 394                 if (sp->hdr.flags & RXRPC_REQUEST_ACK)
 395                         ack_reason = RXRPC_ACK_REQUESTED;
 396                 /* Send an immediate ACK if we fill in a hole */
 397                 else if (!skb_queue_empty(&call->rx_oos_queue))
 398                         ack_reason = RXRPC_ACK_DELAY;
 399                 else
 400                         atomic_inc_return(&call->ackr_nr_unacked);
 401
 402                 window++;
 403                 if (after(window, wtop))
 404                         wtop = window;
 405
 406                 rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg);
 407
 408                 spin_lock(&call->recvmsg_queue.lock);
 409                 rxrpc_input_queue_data(call, skb, window, wtop, rxrpc_receive_queue);
 410                 *_notify = true;
 411
 412                 while ((oos = skb_peek(&call->rx_oos_queue))) {
 413                         struct rxrpc_skb_priv *osp = rxrpc_skb(oos);
 414
 415                         if (after(osp->hdr.seq, window))
 416                                 break;
 417
 418                         __skb_unlink(oos, &call->rx_oos_queue);
 419                         last = osp->hdr.flags & RXRPC_LAST_PACKET;
 420                         seq = osp->hdr.seq;
 421                         if (!reset_sack) {
 422                                 reset_from = seq;
 423                                 reset_sack = true;
 424                         }
 425
 426                         window++;
 427                         rxrpc_input_queue_data(call, oos, window, wtop,
 428                                                  rxrpc_receive_queue_oos);
 429                 }
 430
 431                 spin_unlock(&call->recvmsg_queue.lock);
 432
 433                 if (reset_sack) {
 434                         do {
 435                                 call->ackr_sack_table[reset_from % RXRPC_SACK_SIZE] = 0;
 436                         } while (reset_from++, before(reset_from, window));
 437                 }
 438         } else {
 439                 bool keep = false;
 440
 441                 ack_reason = RXRPC_ACK_OUT_OF_SEQUENCE;
 442
 443                 if (!call->ackr_sack_table[seq % RXRPC_SACK_SIZE]) {
 444                         call->ackr_sack_table[seq % RXRPC_SACK_SIZE] = 1;
 445                         keep = 1;
 446                 }
 447
 448                 if (after(seq + 1, wtop)) {
 449                         wtop = seq + 1;
 450                         rxrpc_input_update_ack_window(call, window, wtop);
 451                 }
 452
 453                 if (!keep) {
 454                         ack_reason = RXRPC_ACK_DUPLICATE;
 455                         goto send_ack;
 456                 }
 457
 458                 skb_queue_walk(&call->rx_oos_queue, oos) {
 459                         struct rxrpc_skb_priv *osp = rxrpc_skb(oos);
 460
 461                         if (after(osp->hdr.seq, seq)) {
 462                                 rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg_oos);
 463                                 __skb_queue_before(&call->rx_oos_queue, oos, skb);
 464                                 goto oos_queued;
 465                         }
 466                 }
 467
 468                 rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg_oos);
 469                 __skb_queue_tail(&call->rx_oos_queue, skb);
 470         oos_queued:
 471                 trace_rxrpc_receive(call, last ? rxrpc_receive_oos_last : rxrpc_receive_oos,
 472                                     sp->hdr.serial, sp->hdr.seq);
 473         }
 474
 475 send_ack:
 476         if (ack_reason >= 0)
 477                 rxrpc_send_ACK(call, ack_reason, serial,
 478                                rxrpc_propose_ack_input_data);
 479         else
 480                 rxrpc_propose_delay_ACK(call, serial,
 481                                         rxrpc_propose_ack_input_data);
 482 }
 483
 484 /*
 485  * Split a jumbo packet and file the bits separately.
 486  */
 487 static bool rxrpc_input_split_jumbo(struct rxrpc_call *call, struct sk_buff *skb)
 488 {
 489         struct rxrpc_jumbo_header jhdr;
 490         struct rxrpc_skb_priv *sp = rxrpc_skb(skb), *jsp;
 491         struct sk_buff *jskb;
 492         unsigned int offset = sizeof(struct rxrpc_wire_header);
 493         unsigned int len = skb->len - offset;
 494         bool notify = false;
 495
 496         while (sp->hdr.flags & RXRPC_JUMBO_PACKET) {
 497                 if (len < RXRPC_JUMBO_SUBPKTLEN)
 498                         goto protocol_error;
 499                 if (sp->hdr.flags & RXRPC_LAST_PACKET)
 500                         goto protocol_error;
 501                 if (skb_copy_bits(skb, offset + RXRPC_JUMBO_DATALEN,
 502                                   &jhdr, sizeof(jhdr)) < 0)
 503                         goto protocol_error;
 504
 505                 jskb = skb_clone(skb, GFP_NOFS);
 506                 if (!jskb) {
 507                         kdebug("couldn't clone");
 508                         return false;
 509                 }
 510                 rxrpc_new_skb(jskb, rxrpc_skb_new_jumbo_subpacket);
 511                 jsp = rxrpc_skb(jskb);
 512                 jsp->offset = offset;
 513                 jsp->len = RXRPC_JUMBO_DATALEN;
 514                 rxrpc_input_data_one(call, jskb, &notify);
 515                 rxrpc_free_skb(jskb, rxrpc_skb_put_jumbo_subpacket);
 516
 517                 sp->hdr.flags = jhdr.flags;
 518                 sp->hdr._rsvd = ntohs(jhdr._rsvd);
 519                 sp->hdr.seq++;
 520                 sp->hdr.serial++;
 521                 offset += RXRPC_JUMBO_SUBPKTLEN;
 522                 len -= RXRPC_JUMBO_SUBPKTLEN;
 523         }
 524
 525         sp->offset = offset;
 526         sp->len    = len;
 527         rxrpc_input_data_one(call, skb, &notify);
 528         if (notify) {
 529                 trace_rxrpc_notify_socket(call->debug_id, sp->hdr.serial);
 530                 rxrpc_notify_socket(call);
 531         }
 532         return true;
 533
 534 protocol_error:
 535         return false;
 536 }
 537
 538 /*
 539  * Process a DATA packet, adding the packet to the Rx ring.  The caller's
 540  * packet ref must be passed on or discarded.
 541  */
 542 static void rxrpc_input_data(struct rxrpc_call *call, struct sk_buff *skb)
 543 {
 544         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 545         enum rxrpc_call_state state;
 546         rxrpc_serial_t serial = sp->hdr.serial;
 547         rxrpc_seq_t seq0 = sp->hdr.seq;
 548
 549         _enter("{%llx,%x},{%u,%x}",
 550                atomic64_read(&call->ackr_window), call->rx_highest_seq,
 551                skb->len, seq0);
 552
 553         state = READ_ONCE(call->state);
 554         if (state >= RXRPC_CALL_COMPLETE)
 555                 return;
 556
 557         if (state == RXRPC_CALL_SERVER_RECV_REQUEST) {
 558                 unsigned long timo = READ_ONCE(call->next_req_timo);
 559                 unsigned long now, expect_req_by;
 560
 561                 if (timo) {
 562                         now = jiffies;
 563                         expect_req_by = now + timo;
 564                         WRITE_ONCE(call->expect_req_by, expect_req_by);
 565                         rxrpc_reduce_call_timer(call, expect_req_by, now,
 566                                                 rxrpc_timer_set_for_idle);
 567                 }
 568         }
 569
 570         /* Received data implicitly ACKs all of the request packets we sent
 571          * when we're acting as a client.
 572          */
 573         if ((state == RXRPC_CALL_CLIENT_SEND_REQUEST ||
 574              state == RXRPC_CALL_CLIENT_AWAIT_REPLY) &&
 575             !rxrpc_receiving_reply(call))
 576                 goto out_notify;
 577
 578         if (!rxrpc_input_split_jumbo(call, skb)) {
 579                 rxrpc_proto_abort("VLD", call, sp->hdr.seq);
 580                 goto out_notify;
 581         }
 582         skb = NULL;
 583
 584 out_notify:
 585         trace_rxrpc_notify_socket(call->debug_id, serial);
 586         rxrpc_notify_socket(call);
 587         _leave(" [queued]");
 588 }
 589
 590 /*
 591  * See if there's a cached RTT probe to complete.
 592  */
 593 static void rxrpc_complete_rtt_probe(struct rxrpc_call *call,
 594                                      ktime_t resp_time,
 595                                      rxrpc_serial_t acked_serial,
 596                                      rxrpc_serial_t ack_serial,
 597                                      enum rxrpc_rtt_rx_trace type)
 598 {
 599         rxrpc_serial_t orig_serial;
 600         unsigned long avail;
 601         ktime_t sent_at;
 602         bool matched = false;
 603         int i;
 604
 605         avail = READ_ONCE(call->rtt_avail);
 606         smp_rmb(); /* Read avail bits before accessing data. */
 607
 608         for (i = 0; i < ARRAY_SIZE(call->rtt_serial); i++) {
 609                 if (!test_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &avail))
 610                         continue;
 611
 612                 sent_at = call->rtt_sent_at[i];
 613                 orig_serial = call->rtt_serial[i];
 614
 615                 if (orig_serial == acked_serial) {
 616                         clear_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
 617                         smp_mb(); /* Read data before setting avail bit */
 618                         set_bit(i, &call->rtt_avail);
 619                         if (type != rxrpc_rtt_rx_cancel)
 620                                 rxrpc_peer_add_rtt(call, type, i, acked_serial, ack_serial,
 621                                                    sent_at, resp_time);
 622                         else
 623                                 trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_cancel, i,
 624                                                    orig_serial, acked_serial, 0, 0);
 625                         matched = true;
 626                 }
 627
 628                 /* If a later serial is being acked, then mark this slot as
 629                  * being available.
 630                  */
 631                 if (after(acked_serial, orig_serial)) {
 632                         trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_obsolete, i,
 633                                            orig_serial, acked_serial, 0, 0);
 634                         clear_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
 635                         smp_wmb();
 636                         set_bit(i, &call->rtt_avail);
 637                 }
 638         }
 639
 640         if (!matched)
 641                 trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_lost, 9, 0, acked_serial, 0, 0);
 642 }
 643
 644 /*
 645  * Process the extra information that may be appended to an ACK packet
 646  */
 647 static void rxrpc_input_ackinfo(struct rxrpc_call *call, struct sk_buff *skb,
 648                                 struct rxrpc_ackinfo *ackinfo)
 649 {
 650         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 651         struct rxrpc_peer *peer;
 652         unsigned int mtu;
 653         bool wake = false;
 654         u32 rwind = ntohl(ackinfo->rwind);
 655
 656         if (rwind > RXRPC_TX_MAX_WINDOW)
 657                 rwind = RXRPC_TX_MAX_WINDOW;
 658         if (call->tx_winsize != rwind) {
 659                 if (rwind > call->tx_winsize)
 660                         wake = true;
 661                 trace_rxrpc_rx_rwind_change(call, sp->hdr.serial, rwind, wake);
 662                 call->tx_winsize = rwind;
 663         }
 664
 665         if (call->cong_ssthresh > rwind)
 666                 call->cong_ssthresh = rwind;
 667
 668         mtu = min(ntohl(ackinfo->rxMTU), ntohl(ackinfo->maxMTU));
 669
 670         peer = call->peer;
 671         if (mtu < peer->maxdata) {
 672                 spin_lock_bh(&peer->lock);
 673                 peer->maxdata = mtu;
 674                 peer->mtu = mtu + peer->hdrsize;
 675                 spin_unlock_bh(&peer->lock);
 676         }
 677
 678         if (wake)
 679                 wake_up(&call->waitq);
 680 }
 681
 682 /*
 683  * Process individual soft ACKs.
 684  *
 685  * Each ACK in the array corresponds to one packet and can be either an ACK or
 686  * a NAK.  If we get find an explicitly NAK'd packet we resend immediately;
 687  * packets that lie beyond the end of the ACK list are scheduled for resend by
 688  * the timer on the basis that the peer might just not have processed them at
 689  * the time the ACK was sent.
 690  */
 691 static void rxrpc_input_soft_acks(struct rxrpc_call *call, u8 *acks,
 692                                   rxrpc_seq_t seq, int nr_acks,
 693                                   struct rxrpc_ack_summary *summary)
 694 {
 695         unsigned int i;
 696
 697         for (i = 0; i < nr_acks; i++) {
 698                 if (acks[i] == RXRPC_ACK_TYPE_ACK) {
 699                         summary->nr_acks++;
 700                         summary->nr_new_acks++;
 701                 } else {
 702                         if (!summary->saw_nacks &&
 703                             call->acks_lowest_nak != seq + i) {
 704                                 call->acks_lowest_nak = seq + i;
 705                                 summary->new_low_nack = true;
 706                         }
 707                         summary->saw_nacks = true;
 708                 }
 709         }
 710 }
 711
 712 /*
 713  * Return true if the ACK is valid - ie. it doesn't appear to have regressed
 714  * with respect to the ack state conveyed by preceding ACKs.
 715  */
 716 static bool rxrpc_is_ack_valid(struct rxrpc_call *call,
 717                                rxrpc_seq_t first_pkt, rxrpc_seq_t prev_pkt)
 718 {
 719         rxrpc_seq_t base = READ_ONCE(call->acks_first_seq);
 720
 721         if (after(first_pkt, base))
 722                 return true; /* The window advanced */
 723
 724         if (before(first_pkt, base))
 725                 return false; /* firstPacket regressed */
 726
 727         if (after_eq(prev_pkt, call->acks_prev_seq))
 728                 return true; /* previousPacket hasn't regressed. */
 729
 730         /* Some rx implementations put a serial number in previousPacket. */
 731         if (after_eq(prev_pkt, base + call->tx_winsize))
 732                 return false;
 733         return true;
 734 }
 735
 736 /*
 737  * Process an ACK packet.
 738  *
 739  * ack.firstPacket is the sequence number of the first soft-ACK'd/NAK'd packet
 740  * in the ACK array.  Anything before that is hard-ACK'd and may be discarded.
 741  *
 742  * A hard-ACK means that a packet has been processed and may be discarded; a
 743  * soft-ACK means that the packet may be discarded and retransmission
 744  * requested.  A phase is complete when all packets are hard-ACK'd.
 745  */
 746 static void rxrpc_input_ack(struct rxrpc_call *call, struct sk_buff *skb)
 747 {
 748         struct rxrpc_ack_summary summary = { 0 };
 749         struct rxrpc_ackpacket ack;
 750         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 751         struct rxrpc_ackinfo info;
 752         rxrpc_serial_t ack_serial, acked_serial;
 753         rxrpc_seq_t first_soft_ack, hard_ack, prev_pkt;
 754         int nr_acks, offset, ioffset;
 755
 756         _enter("");
 757
 758         offset = sizeof(struct rxrpc_wire_header);
 759         if (skb_copy_bits(skb, offset, &ack, sizeof(ack)) < 0)
 760                 return rxrpc_proto_abort("XAK", call, 0);
 761         offset += sizeof(ack);
 762
 763         ack_serial = sp->hdr.serial;
 764         acked_serial = ntohl(ack.serial);
 765         first_soft_ack = ntohl(ack.firstPacket);
 766         prev_pkt = ntohl(ack.previousPacket);
 767         hard_ack = first_soft_ack - 1;
 768         nr_acks = ack.nAcks;
 769         summary.ack_reason = (ack.reason < RXRPC_ACK__INVALID ?
 770                               ack.reason : RXRPC_ACK__INVALID);
 771
 772         trace_rxrpc_rx_ack(call, ack_serial, acked_serial,
 773                            first_soft_ack, prev_pkt,
 774                            summary.ack_reason, nr_acks);
 775         rxrpc_inc_stat(call->rxnet, stat_rx_acks[ack.reason]);
 776
 777         switch (ack.reason) {
 778         case RXRPC_ACK_PING_RESPONSE:
 779                 rxrpc_complete_rtt_probe(call, skb->tstamp, acked_serial, ack_serial,
 780                                          rxrpc_rtt_rx_ping_response);
 781                 break;
 782         case RXRPC_ACK_REQUESTED:
 783                 rxrpc_complete_rtt_probe(call, skb->tstamp, acked_serial, ack_serial,
 784                                          rxrpc_rtt_rx_requested_ack);
 785                 break;
 786         default:
 787                 if (acked_serial != 0)
 788                         rxrpc_complete_rtt_probe(call, skb->tstamp, acked_serial, ack_serial,
 789                                                  rxrpc_rtt_rx_cancel);
 790                 break;
 791         }
 792
 793         if (ack.reason == RXRPC_ACK_PING) {
 794                 rxrpc_send_ACK(call, RXRPC_ACK_PING_RESPONSE, ack_serial,
 795                                rxrpc_propose_ack_respond_to_ping);
 796         } else if (sp->hdr.flags & RXRPC_REQUEST_ACK) {
 797                 rxrpc_send_ACK(call, RXRPC_ACK_REQUESTED, ack_serial,
 798                                rxrpc_propose_ack_respond_to_ack);
 799         }
 800
 801         /* If we get an EXCEEDS_WINDOW ACK from the server, it probably
 802          * indicates that the client address changed due to NAT.  The server
 803          * lost the call because it switched to a different peer.
 804          */
 805         if (unlikely(ack.reason == RXRPC_ACK_EXCEEDS_WINDOW) &&
 806             first_soft_ack == 1 &&
 807             prev_pkt == 0 &&
 808             rxrpc_is_client_call(call)) {
 809                 rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
 810                                           0, -ENETRESET);
 811                 return;
 812         }
 813
 814         /* If we get an OUT_OF_SEQUENCE ACK from the server, that can also
 815          * indicate a change of address.  However, we can retransmit the call
 816          * if we still have it buffered to the beginning.
 817          */
 818         if (unlikely(ack.reason == RXRPC_ACK_OUT_OF_SEQUENCE) &&
 819             first_soft_ack == 1 &&
 820             prev_pkt == 0 &&
 821             call->acks_hard_ack == 0 &&
 822             rxrpc_is_client_call(call)) {
 823                 rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
 824                                           0, -ENETRESET);
 825                 return;
 826         }
 827
 828         /* Discard any out-of-order or duplicate ACKs (outside lock). */
 829         if (!rxrpc_is_ack_valid(call, first_soft_ack, prev_pkt)) {
 830                 trace_rxrpc_rx_discard_ack(call->debug_id, ack_serial,
 831                                            first_soft_ack, call->acks_first_seq,
 832                                            prev_pkt, call->acks_prev_seq);
 833                 return;
 834         }
 835
 836         info.rxMTU = 0;
 837         ioffset = offset + nr_acks + 3;
 838         if (skb->len >= ioffset + sizeof(info) &&
 839             skb_copy_bits(skb, ioffset, &info, sizeof(info)) < 0)
 840                 return rxrpc_proto_abort("XAI", call, 0);
 841
 842         if (nr_acks > 0)
 843                 skb_condense(skb);
 844
 845         call->acks_latest_ts = skb->tstamp;
 846         call->acks_first_seq = first_soft_ack;
 847         call->acks_prev_seq = prev_pkt;
 848
 849         switch (ack.reason) {
 850         case RXRPC_ACK_PING:
 851                 break;
 852         default:
 853                 if (after(acked_serial, call->acks_highest_serial))
 854                         call->acks_highest_serial = acked_serial;
 855                 break;
 856         }
 857
 858         /* Parse rwind and mtu sizes if provided. */
 859         if (info.rxMTU)
 860                 rxrpc_input_ackinfo(call, skb, &info);
 861
 862         if (first_soft_ack == 0)
 863                 return rxrpc_proto_abort("AK0", call, 0);
 864
 865         /* Ignore ACKs unless we are or have just been transmitting. */
 866         switch (READ_ONCE(call->state)) {
 867         case RXRPC_CALL_CLIENT_SEND_REQUEST:
 868         case RXRPC_CALL_CLIENT_AWAIT_REPLY:
 869         case RXRPC_CALL_SERVER_SEND_REPLY:
 870         case RXRPC_CALL_SERVER_AWAIT_ACK:
 871                 break;
 872         default:
 873                 return;
 874         }
 875
 876         if (before(hard_ack, call->acks_hard_ack) ||
 877             after(hard_ack, call->tx_top))
 878                 return rxrpc_proto_abort("AKW", call, 0);
 879         if (nr_acks > call->tx_top - hard_ack)
 880                 return rxrpc_proto_abort("AKN", call, 0);
 881
 882         if (after(hard_ack, call->acks_hard_ack)) {
 883                 if (rxrpc_rotate_tx_window(call, hard_ack, &summary)) {
 884                         rxrpc_end_tx_phase(call, false, "ETA");
 885                         return;
 886                 }
 887         }
 888
 889         if (nr_acks > 0) {
 890                 if (offset > (int)skb->len - nr_acks)
 891                         return rxrpc_proto_abort("XSA", call, 0);
 892                 rxrpc_input_soft_acks(call, skb->data + offset, first_soft_ack,
 893                                       nr_acks, &summary);
 894         }
 895
 896         if (test_bit(RXRPC_CALL_TX_LAST, &call->flags) &&
 897             summary.nr_acks == call->tx_top - hard_ack &&
 898             rxrpc_is_client_call(call))
 899                 rxrpc_propose_ping(call, ack_serial,
 900                                    rxrpc_propose_ack_ping_for_lost_reply);
 901
 902         rxrpc_congestion_management(call, skb, &summary, acked_serial);
 903 }
 904
 905 /*
 906  * Process an ACKALL packet.
 907  */
 908 static void rxrpc_input_ackall(struct rxrpc_call *call, struct sk_buff *skb)
 909 {
 910         struct rxrpc_ack_summary summary = { 0 };
 911
 912         if (rxrpc_rotate_tx_window(call, call->tx_top, &summary))
 913                 rxrpc_end_tx_phase(call, false, "ETL");
 914 }
 915
 916 /*
 917  * Process an ABORT packet directed at a call.
 918  */
 919 static void rxrpc_input_abort(struct rxrpc_call *call, struct sk_buff *skb)
 920 {
 921         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 922
 923         trace_rxrpc_rx_abort(call, sp->hdr.serial, skb->priority);
 924
 925         rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
 926                                   skb->priority, -ECONNABORTED);
 927 }
 928
 929 /*
 930  * Process an incoming call packet.
 931  */
 932 void rxrpc_input_call_packet(struct rxrpc_call *call, struct sk_buff *skb)
 933 {
 934         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 935         unsigned long timo;
 936
 937         _enter("%p,%p", call, skb);
 938
 939         if (sp->hdr.serviceId != call->dest_srx.srx_service)
 940                 call->dest_srx.srx_service = sp->hdr.serviceId;
 941         if ((int)sp->hdr.serial - (int)call->rx_serial > 0)
 942                 call->rx_serial = sp->hdr.serial;
 943         if (!test_bit(RXRPC_CALL_RX_HEARD, &call->flags))
 944                 set_bit(RXRPC_CALL_RX_HEARD, &call->flags);
 945
 946         timo = READ_ONCE(call->next_rx_timo);
 947         if (timo) {
 948                 unsigned long now = jiffies, expect_rx_by;
 949
 950                 expect_rx_by = now + timo;
 951                 WRITE_ONCE(call->expect_rx_by, expect_rx_by);
 952                 rxrpc_reduce_call_timer(call, expect_rx_by, now,
 953                                         rxrpc_timer_set_for_normal);
 954         }
 955
 956         switch (sp->hdr.type) {
 957         case RXRPC_PACKET_TYPE_DATA:
 958                 rxrpc_input_data(call, skb);
 959                 break;
 960
 961         case RXRPC_PACKET_TYPE_ACK:
 962                 rxrpc_input_ack(call, skb);
 963                 break;
 964
 965         case RXRPC_PACKET_TYPE_BUSY:
 966                 /* Just ignore BUSY packets from the server; the retry and
 967                  * lifespan timers will take care of business.  BUSY packets
 968                  * from the client don't make sense.
 969                  */
 970                 break;
 971
 972         case RXRPC_PACKET_TYPE_ABORT:
 973                 rxrpc_input_abort(call, skb);
 974                 break;
 975
 976         case RXRPC_PACKET_TYPE_ACKALL:
 977                 rxrpc_input_ackall(call, skb);
 978                 break;
 979
 980         default:
 981                 break;
 982         }
 983 }
 984
 985 /*
 986  * Handle a new service call on a channel implicitly completing the preceding
 987  * call on that channel.  This does not apply to client conns.
 988  *
 989  * TODO: If callNumber > call_id + 1, renegotiate security.
 990  */
 991 void rxrpc_implicit_end_call(struct rxrpc_call *call, struct sk_buff *skb)
 992 {
 993         struct rxrpc_connection *conn = call->conn;
 994
 995         switch (READ_ONCE(call->state)) {
 996         case RXRPC_CALL_SERVER_AWAIT_ACK:
 997                 rxrpc_call_completed(call);
 998                 fallthrough;
 999         case RXRPC_CALL_COMPLETE:
1000                 break;
1001         default:
1002                 if (rxrpc_abort_call("IMP", call, 0, RX_CALL_DEAD, -ESHUTDOWN))
1003                         rxrpc_send_abort_packet(call);
1004                 trace_rxrpc_improper_term(call);
1005                 break;
1006         }
1007
1008         rxrpc_input_call_event(call, skb);
1009
1010         spin_lock(&conn->bundle->channel_lock);
1011         __rxrpc_disconnect_call(conn, call);
1012         spin_unlock(&conn->bundle->channel_lock);
1013 }