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         rxrpc_abort_call(why, call, seq, RX_PROTOCOL_ERROR, -EBADMSG);
  16 }
  17
  18 /*
  19  * Do TCP-style congestion management [RFC 5681].
  20  */
  21 static void rxrpc_congestion_management(struct rxrpc_call *call,
  22                                         struct sk_buff *skb,
  23                                         struct rxrpc_ack_summary *summary,
  24                                         rxrpc_serial_t acked_serial)
  25 {
  26         enum rxrpc_congest_change change = rxrpc_cong_no_change;
  27         unsigned int cumulative_acks = call->cong_cumul_acks;
  28         unsigned int cwnd = call->cong_cwnd;
  29         bool resend = false;
  30
  31         summary->flight_size =
  32                 (call->tx_top - call->acks_hard_ack) - summary->nr_acks;
  33
  34         if (test_and_clear_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags)) {
  35                 summary->retrans_timeo = true;
  36                 call->cong_ssthresh = max_t(unsigned int,
  37                                             summary->flight_size / 2, 2);
  38                 cwnd = 1;
  39                 if (cwnd >= call->cong_ssthresh &&
  40                     call->cong_mode == RXRPC_CALL_SLOW_START) {
  41                         call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
  42                         call->cong_tstamp = skb->tstamp;
  43                         cumulative_acks = 0;
  44                 }
  45         }
  46
  47         cumulative_acks += summary->nr_new_acks;
  48         cumulative_acks += summary->nr_rot_new_acks;
  49         if (cumulative_acks > 255)
  50                 cumulative_acks = 255;
  51
  52         summary->mode = call->cong_mode;
  53         summary->cwnd = call->cong_cwnd;
  54         summary->ssthresh = call->cong_ssthresh;
  55         summary->cumulative_acks = cumulative_acks;
  56         summary->dup_acks = call->cong_dup_acks;
  57
  58         switch (call->cong_mode) {
  59         case RXRPC_CALL_SLOW_START:
  60                 if (summary->saw_nacks)
  61                         goto packet_loss_detected;
  62                 if (summary->cumulative_acks > 0)
  63                         cwnd += 1;
  64                 if (cwnd >= call->cong_ssthresh) {
  65                         call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
  66                         call->cong_tstamp = skb->tstamp;
  67                 }
  68                 goto out;
  69
  70         case RXRPC_CALL_CONGEST_AVOIDANCE:
  71                 if (summary->saw_nacks)
  72                         goto packet_loss_detected;
  73
  74                 /* We analyse the number of packets that get ACK'd per RTT
  75                  * period and increase the window if we managed to fill it.
  76                  */
  77                 if (call->peer->rtt_count == 0)
  78                         goto out;
  79                 if (ktime_before(skb->tstamp,
  80                                  ktime_add_us(call->cong_tstamp,
  81                                               call->peer->srtt_us >> 3)))
  82                         goto out_no_clear_ca;
  83                 change = rxrpc_cong_rtt_window_end;
  84                 call->cong_tstamp = skb->tstamp;
  85                 if (cumulative_acks >= cwnd)
  86                         cwnd++;
  87                 goto out;
  88
  89         case RXRPC_CALL_PACKET_LOSS:
  90                 if (!summary->saw_nacks)
  91                         goto resume_normality;
  92
  93                 if (summary->new_low_nack) {
  94                         change = rxrpc_cong_new_low_nack;
  95                         call->cong_dup_acks = 1;
  96                         if (call->cong_extra > 1)
  97                                 call->cong_extra = 1;
  98                         goto send_extra_data;
  99                 }
 100
 101                 call->cong_dup_acks++;
 102                 if (call->cong_dup_acks < 3)
 103                         goto send_extra_data;
 104
 105                 change = rxrpc_cong_begin_retransmission;
 106                 call->cong_mode = RXRPC_CALL_FAST_RETRANSMIT;
 107                 call->cong_ssthresh = max_t(unsigned int,
 108                                             summary->flight_size / 2, 2);
 109                 cwnd = call->cong_ssthresh + 3;
 110                 call->cong_extra = 0;
 111                 call->cong_dup_acks = 0;
 112                 resend = true;
 113                 goto out;
 114
 115         case RXRPC_CALL_FAST_RETRANSMIT:
 116                 if (!summary->new_low_nack) {
 117                         if (summary->nr_new_acks == 0)
 118                                 cwnd += 1;
 119                         call->cong_dup_acks++;
 120                         if (call->cong_dup_acks == 2) {
 121                                 change = rxrpc_cong_retransmit_again;
 122                                 call->cong_dup_acks = 0;
 123                                 resend = true;
 124                         }
 125                 } else {
 126                         change = rxrpc_cong_progress;
 127                         cwnd = call->cong_ssthresh;
 128                         if (!summary->saw_nacks)
 129                                 goto resume_normality;
 130                 }
 131                 goto out;
 132
 133         default:
 134                 BUG();
 135                 goto out;
 136         }
 137
 138 resume_normality:
 139         change = rxrpc_cong_cleared_nacks;
 140         call->cong_dup_acks = 0;
 141         call->cong_extra = 0;
 142         call->cong_tstamp = skb->tstamp;
 143         if (cwnd < call->cong_ssthresh)
 144                 call->cong_mode = RXRPC_CALL_SLOW_START;
 145         else
 146                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
 147 out:
 148         cumulative_acks = 0;
 149 out_no_clear_ca:
 150         if (cwnd >= RXRPC_TX_MAX_WINDOW)
 151                 cwnd = RXRPC_TX_MAX_WINDOW;
 152         call->cong_cwnd = cwnd;
 153         call->cong_cumul_acks = cumulative_acks;
 154         trace_rxrpc_congest(call, summary, acked_serial, change);
 155         if (resend)
 156                 rxrpc_resend(call, skb);
 157         return;
 158
 159 packet_loss_detected:
 160         change = rxrpc_cong_saw_nack;
 161         call->cong_mode = RXRPC_CALL_PACKET_LOSS;
 162         call->cong_dup_acks = 0;
 163         goto send_extra_data;
 164
 165 send_extra_data:
 166         /* Send some previously unsent DATA if we have some to advance the ACK
 167          * state.
 168          */
 169         if (test_bit(RXRPC_CALL_TX_LAST, &call->flags) ||
 170             summary->nr_acks != call->tx_top - call->acks_hard_ack) {
 171                 call->cong_extra++;
 172                 wake_up(&call->waitq);
 173         }
 174         goto out_no_clear_ca;
 175 }
 176
 177 /*
 178  * Degrade the congestion window if we haven't transmitted a packet for >1RTT.
 179  */
 180 void rxrpc_congestion_degrade(struct rxrpc_call *call)
 181 {
 182         ktime_t rtt, now;
 183
 184         if (call->cong_mode != RXRPC_CALL_SLOW_START &&
 185             call->cong_mode != RXRPC_CALL_CONGEST_AVOIDANCE)
 186                 return;
 187         if (call->state == RXRPC_CALL_CLIENT_AWAIT_REPLY)
 188                 return;
 189
 190         rtt = ns_to_ktime(call->peer->srtt_us * (1000 / 8));
 191         now = ktime_get_real();
 192         if (!ktime_before(ktime_add(call->tx_last_sent, rtt), now))
 193                 return;
 194
 195         trace_rxrpc_reset_cwnd(call, now);
 196         rxrpc_inc_stat(call->rxnet, stat_tx_data_cwnd_reset);
 197         call->tx_last_sent = now;
 198         call->cong_mode = RXRPC_CALL_SLOW_START;
 199         call->cong_ssthresh = max_t(unsigned int, call->cong_ssthresh,
 200                                     call->cong_cwnd * 3 / 4);
 201         call->cong_cwnd = max_t(unsigned int, call->cong_cwnd / 2, RXRPC_MIN_CWND);
 202 }
 203
 204 /*
 205  * Apply a hard ACK by advancing the Tx window.
 206  */
 207 static bool rxrpc_rotate_tx_window(struct rxrpc_call *call, rxrpc_seq_t to,
 208                                    struct rxrpc_ack_summary *summary)
 209 {
 210         struct rxrpc_txbuf *txb;
 211         bool rot_last = false;
 212
 213         list_for_each_entry_rcu(txb, &call->tx_buffer, call_link, false) {
 214                 if (before_eq(txb->seq, call->acks_hard_ack))
 215                         continue;
 216                 summary->nr_rot_new_acks++;
 217                 if (test_bit(RXRPC_TXBUF_LAST, &txb->flags)) {
 218                         set_bit(RXRPC_CALL_TX_LAST, &call->flags);
 219                         rot_last = true;
 220                 }
 221                 if (txb->seq == to)
 222                         break;
 223         }
 224
 225         if (rot_last)
 226                 set_bit(RXRPC_CALL_TX_ALL_ACKED, &call->flags);
 227
 228         _enter("%x,%x,%x,%d", to, call->acks_hard_ack, call->tx_top, rot_last);
 229
 230         if (call->acks_lowest_nak == call->acks_hard_ack) {
 231                 call->acks_lowest_nak = to;
 232         } else if (after(to, call->acks_lowest_nak)) {
 233                 summary->new_low_nack = true;
 234                 call->acks_lowest_nak = to;
 235         }
 236
 237         smp_store_release(&call->acks_hard_ack, to);
 238
 239         trace_rxrpc_txqueue(call, (rot_last ?
 240                                    rxrpc_txqueue_rotate_last :
 241                                    rxrpc_txqueue_rotate));
 242         wake_up(&call->waitq);
 243         return rot_last;
 244 }
 245
 246 /*
 247  * End the transmission phase of a call.
 248  *
 249  * This occurs when we get an ACKALL packet, the first DATA packet of a reply,
 250  * or a final ACK packet.
 251  */
 252 static bool rxrpc_end_tx_phase(struct rxrpc_call *call, bool reply_begun,
 253                                const char *abort_why)
 254 {
 255         unsigned int state;
 256
 257         ASSERT(test_bit(RXRPC_CALL_TX_LAST, &call->flags));
 258
 259         write_lock(&call->state_lock);
 260
 261         state = call->state;
 262         switch (state) {
 263         case RXRPC_CALL_CLIENT_SEND_REQUEST:
 264         case RXRPC_CALL_CLIENT_AWAIT_REPLY:
 265                 if (reply_begun)
 266                         call->state = state = RXRPC_CALL_CLIENT_RECV_REPLY;
 267                 else
 268                         call->state = state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
 269                 break;
 270
 271         case RXRPC_CALL_SERVER_AWAIT_ACK:
 272                 __rxrpc_call_completed(call);
 273                 state = call->state;
 274                 break;
 275
 276         default:
 277                 goto bad_state;
 278         }
 279
 280         write_unlock(&call->state_lock);
 281         if (state == RXRPC_CALL_CLIENT_AWAIT_REPLY)
 282                 trace_rxrpc_txqueue(call, rxrpc_txqueue_await_reply);
 283         else
 284                 trace_rxrpc_txqueue(call, rxrpc_txqueue_end);
 285         _leave(" = ok");
 286         return true;
 287
 288 bad_state:
 289         write_unlock(&call->state_lock);
 290         kdebug("end_tx %s", rxrpc_call_states[call->state]);
 291         rxrpc_proto_abort(abort_why, call, call->tx_top);
 292         return false;
 293 }
 294
 295 /*
 296  * Begin the reply reception phase of a call.
 297  */
 298 static bool rxrpc_receiving_reply(struct rxrpc_call *call)
 299 {
 300         struct rxrpc_ack_summary summary = { 0 };
 301         unsigned long now, timo;
 302         rxrpc_seq_t top = READ_ONCE(call->tx_top);
 303
 304         if (call->ackr_reason) {
 305                 now = jiffies;
 306                 timo = now + MAX_JIFFY_OFFSET;
 307                 WRITE_ONCE(call->resend_at, timo);
 308                 WRITE_ONCE(call->delay_ack_at, timo);
 309                 trace_rxrpc_timer(call, rxrpc_timer_init_for_reply, now);
 310         }
 311
 312         if (!test_bit(RXRPC_CALL_TX_LAST, &call->flags)) {
 313                 if (!rxrpc_rotate_tx_window(call, top, &summary)) {
 314                         rxrpc_proto_abort("TXL", call, top);
 315                         return false;
 316                 }
 317         }
 318         return rxrpc_end_tx_phase(call, true, "ETD");
 319 }
 320
 321 static void rxrpc_input_update_ack_window(struct rxrpc_call *call,
 322                                           rxrpc_seq_t window, rxrpc_seq_t wtop)
 323 {
 324         atomic64_set_release(&call->ackr_window, ((u64)wtop) << 32 | window);
 325 }
 326
 327 /*
 328  * Push a DATA packet onto the Rx queue.
 329  */
 330 static void rxrpc_input_queue_data(struct rxrpc_call *call, struct sk_buff *skb,
 331                                    rxrpc_seq_t window, rxrpc_seq_t wtop,
 332                                    enum rxrpc_receive_trace why)
 333 {
 334         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 335         bool last = sp->hdr.flags & RXRPC_LAST_PACKET;
 336
 337         __skb_queue_tail(&call->recvmsg_queue, skb);
 338         rxrpc_input_update_ack_window(call, window, wtop);
 339
 340         trace_rxrpc_receive(call, last ? why + 1 : why, sp->hdr.serial, sp->hdr.seq);
 341 }
 342
 343 /*
 344  * Process a DATA packet.
 345  */
 346 static void rxrpc_input_data_one(struct rxrpc_call *call, struct sk_buff *skb,
 347                                  bool *_notify)
 348 {
 349         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 350         struct sk_buff *oos;
 351         rxrpc_serial_t serial = sp->hdr.serial;
 352         u64 win = atomic64_read(&call->ackr_window);
 353         rxrpc_seq_t window = lower_32_bits(win);
 354         rxrpc_seq_t wtop = upper_32_bits(win);
 355         rxrpc_seq_t wlimit = window + call->rx_winsize - 1;
 356         rxrpc_seq_t seq = sp->hdr.seq;
 357         bool last = sp->hdr.flags & RXRPC_LAST_PACKET;
 358         int ack_reason = -1;
 359
 360         rxrpc_inc_stat(call->rxnet, stat_rx_data);
 361         if (sp->hdr.flags & RXRPC_REQUEST_ACK)
 362                 rxrpc_inc_stat(call->rxnet, stat_rx_data_reqack);
 363         if (sp->hdr.flags & RXRPC_JUMBO_PACKET)
 364                 rxrpc_inc_stat(call->rxnet, stat_rx_data_jumbo);
 365
 366         if (last) {
 367                 if (test_and_set_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
 368                     seq + 1 != wtop) {
 369                         rxrpc_proto_abort("LSN", call, seq);
 370                         return;
 371                 }
 372         } else {
 373                 if (test_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
 374                     after_eq(seq, wtop)) {
 375                         pr_warn("Packet beyond last: c=%x q=%x window=%x-%x wlimit=%x\n",
 376                                 call->debug_id, seq, window, wtop, wlimit);
 377                         rxrpc_proto_abort("LSA", call, seq);
 378                         return;
 379                 }
 380         }
 381
 382         if (after(seq, call->rx_highest_seq))
 383                 call->rx_highest_seq = seq;
 384
 385         trace_rxrpc_rx_data(call->debug_id, seq, serial, sp->hdr.flags);
 386
 387         if (before(seq, window)) {
 388                 ack_reason = RXRPC_ACK_DUPLICATE;
 389                 goto send_ack;
 390         }
 391         if (after(seq, wlimit)) {
 392                 ack_reason = RXRPC_ACK_EXCEEDS_WINDOW;
 393                 goto send_ack;
 394         }
 395
 396         /* Queue the packet. */
 397         if (seq == window) {
 398                 rxrpc_seq_t reset_from;
 399                 bool reset_sack = false;
 400
 401                 if (sp->hdr.flags & RXRPC_REQUEST_ACK)
 402                         ack_reason = RXRPC_ACK_REQUESTED;
 403                 /* Send an immediate ACK if we fill in a hole */
 404                 else if (!skb_queue_empty(&call->rx_oos_queue))
 405                         ack_reason = RXRPC_ACK_DELAY;
 406                 else
 407                         atomic_inc_return(&call->ackr_nr_unacked);
 408
 409                 window++;
 410                 if (after(window, wtop))
 411                         wtop = window;
 412
 413                 rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg);
 414
 415                 spin_lock(&call->recvmsg_queue.lock);
 416                 rxrpc_input_queue_data(call, skb, window, wtop, rxrpc_receive_queue);
 417                 *_notify = true;
 418
 419                 while ((oos = skb_peek(&call->rx_oos_queue))) {
 420                         struct rxrpc_skb_priv *osp = rxrpc_skb(oos);
 421
 422                         if (after(osp->hdr.seq, window))
 423                                 break;
 424
 425                         __skb_unlink(oos, &call->rx_oos_queue);
 426                         last = osp->hdr.flags & RXRPC_LAST_PACKET;
 427                         seq = osp->hdr.seq;
 428                         if (!reset_sack) {
 429                                 reset_from = seq;
 430                                 reset_sack = true;
 431                         }
 432
 433                         window++;
 434                         rxrpc_input_queue_data(call, oos, window, wtop,
 435                                                  rxrpc_receive_queue_oos);
 436                 }
 437
 438                 spin_unlock(&call->recvmsg_queue.lock);
 439
 440                 if (reset_sack) {
 441                         do {
 442                                 call->ackr_sack_table[reset_from % RXRPC_SACK_SIZE] = 0;
 443                         } while (reset_from++, before(reset_from, window));
 444                 }
 445         } else {
 446                 bool keep = false;
 447
 448                 ack_reason = RXRPC_ACK_OUT_OF_SEQUENCE;
 449
 450                 if (!call->ackr_sack_table[seq % RXRPC_SACK_SIZE]) {
 451                         call->ackr_sack_table[seq % RXRPC_SACK_SIZE] = 1;
 452                         keep = 1;
 453                 }
 454
 455                 if (after(seq + 1, wtop)) {
 456                         wtop = seq + 1;
 457                         rxrpc_input_update_ack_window(call, window, wtop);
 458                 }
 459
 460                 if (!keep) {
 461                         ack_reason = RXRPC_ACK_DUPLICATE;
 462                         goto send_ack;
 463                 }
 464
 465                 skb_queue_walk(&call->rx_oos_queue, oos) {
 466                         struct rxrpc_skb_priv *osp = rxrpc_skb(oos);
 467
 468                         if (after(osp->hdr.seq, seq)) {
 469                                 rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg_oos);
 470                                 __skb_queue_before(&call->rx_oos_queue, oos, skb);
 471                                 goto oos_queued;
 472                         }
 473                 }
 474
 475                 rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg_oos);
 476                 __skb_queue_tail(&call->rx_oos_queue, skb);
 477         oos_queued:
 478                 trace_rxrpc_receive(call, last ? rxrpc_receive_oos_last : rxrpc_receive_oos,
 479                                     sp->hdr.serial, sp->hdr.seq);
 480         }
 481
 482 send_ack:
 483         if (ack_reason >= 0)
 484                 rxrpc_send_ACK(call, ack_reason, serial,
 485                                rxrpc_propose_ack_input_data);
 486         else
 487                 rxrpc_propose_delay_ACK(call, serial,
 488                                         rxrpc_propose_ack_input_data);
 489 }
 490
 491 /*
 492  * Split a jumbo packet and file the bits separately.
 493  */
 494 static bool rxrpc_input_split_jumbo(struct rxrpc_call *call, struct sk_buff *skb)
 495 {
 496         struct rxrpc_jumbo_header jhdr;
 497         struct rxrpc_skb_priv *sp = rxrpc_skb(skb), *jsp;
 498         struct sk_buff *jskb;
 499         unsigned int offset = sizeof(struct rxrpc_wire_header);
 500         unsigned int len = skb->len - offset;
 501         bool notify = false;
 502
 503         while (sp->hdr.flags & RXRPC_JUMBO_PACKET) {
 504                 if (len < RXRPC_JUMBO_SUBPKTLEN)
 505                         goto protocol_error;
 506                 if (sp->hdr.flags & RXRPC_LAST_PACKET)
 507                         goto protocol_error;
 508                 if (skb_copy_bits(skb, offset + RXRPC_JUMBO_DATALEN,
 509                                   &jhdr, sizeof(jhdr)) < 0)
 510                         goto protocol_error;
 511
 512                 jskb = skb_clone(skb, GFP_NOFS);
 513                 if (!jskb) {
 514                         kdebug("couldn't clone");
 515                         return false;
 516                 }
 517                 rxrpc_new_skb(jskb, rxrpc_skb_new_jumbo_subpacket);
 518                 jsp = rxrpc_skb(jskb);
 519                 jsp->offset = offset;
 520                 jsp->len = RXRPC_JUMBO_DATALEN;
 521                 rxrpc_input_data_one(call, jskb, &notify);
 522                 rxrpc_free_skb(jskb, rxrpc_skb_put_jumbo_subpacket);
 523
 524                 sp->hdr.flags = jhdr.flags;
 525                 sp->hdr._rsvd = ntohs(jhdr._rsvd);
 526                 sp->hdr.seq++;
 527                 sp->hdr.serial++;
 528                 offset += RXRPC_JUMBO_SUBPKTLEN;
 529                 len -= RXRPC_JUMBO_SUBPKTLEN;
 530         }
 531
 532         sp->offset = offset;
 533         sp->len    = len;
 534         rxrpc_input_data_one(call, skb, &notify);
 535         if (notify) {
 536                 trace_rxrpc_notify_socket(call->debug_id, sp->hdr.serial);
 537                 rxrpc_notify_socket(call);
 538         }
 539         return true;
 540
 541 protocol_error:
 542         return false;
 543 }
 544
 545 /*
 546  * Process a DATA packet, adding the packet to the Rx ring.  The caller's
 547  * packet ref must be passed on or discarded.
 548  */
 549 static void rxrpc_input_data(struct rxrpc_call *call, struct sk_buff *skb)
 550 {
 551         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 552         enum rxrpc_call_state state;
 553         rxrpc_serial_t serial = sp->hdr.serial;
 554         rxrpc_seq_t seq0 = sp->hdr.seq;
 555
 556         _enter("{%llx,%x},{%u,%x}",
 557                atomic64_read(&call->ackr_window), call->rx_highest_seq,
 558                skb->len, seq0);
 559
 560         state = READ_ONCE(call->state);
 561         if (state >= RXRPC_CALL_COMPLETE)
 562                 return;
 563
 564         if (state == RXRPC_CALL_SERVER_RECV_REQUEST) {
 565                 unsigned long timo = READ_ONCE(call->next_req_timo);
 566                 unsigned long now, expect_req_by;
 567
 568                 if (timo) {
 569                         now = jiffies;
 570                         expect_req_by = now + timo;
 571                         WRITE_ONCE(call->expect_req_by, expect_req_by);
 572                         rxrpc_reduce_call_timer(call, expect_req_by, now,
 573                                                 rxrpc_timer_set_for_idle);
 574                 }
 575         }
 576
 577         /* Received data implicitly ACKs all of the request packets we sent
 578          * when we're acting as a client.
 579          */
 580         if ((state == RXRPC_CALL_CLIENT_SEND_REQUEST ||
 581              state == RXRPC_CALL_CLIENT_AWAIT_REPLY) &&
 582             !rxrpc_receiving_reply(call))
 583                 goto out_notify;
 584
 585         if (!rxrpc_input_split_jumbo(call, skb)) {
 586                 rxrpc_proto_abort("VLD", call, sp->hdr.seq);
 587                 goto out_notify;
 588         }
 589         skb = NULL;
 590
 591 out_notify:
 592         trace_rxrpc_notify_socket(call->debug_id, serial);
 593         rxrpc_notify_socket(call);
 594         _leave(" [queued]");
 595 }
 596
 597 /*
 598  * See if there's a cached RTT probe to complete.
 599  */
 600 static void rxrpc_complete_rtt_probe(struct rxrpc_call *call,
 601                                      ktime_t resp_time,
 602                                      rxrpc_serial_t acked_serial,
 603                                      rxrpc_serial_t ack_serial,
 604                                      enum rxrpc_rtt_rx_trace type)
 605 {
 606         rxrpc_serial_t orig_serial;
 607         unsigned long avail;
 608         ktime_t sent_at;
 609         bool matched = false;
 610         int i;
 611
 612         avail = READ_ONCE(call->rtt_avail);
 613         smp_rmb(); /* Read avail bits before accessing data. */
 614
 615         for (i = 0; i < ARRAY_SIZE(call->rtt_serial); i++) {
 616                 if (!test_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &avail))
 617                         continue;
 618
 619                 sent_at = call->rtt_sent_at[i];
 620                 orig_serial = call->rtt_serial[i];
 621
 622                 if (orig_serial == acked_serial) {
 623                         clear_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
 624                         smp_mb(); /* Read data before setting avail bit */
 625                         set_bit(i, &call->rtt_avail);
 626                         if (type != rxrpc_rtt_rx_cancel)
 627                                 rxrpc_peer_add_rtt(call, type, i, acked_serial, ack_serial,
 628                                                    sent_at, resp_time);
 629                         else
 630                                 trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_cancel, i,
 631                                                    orig_serial, acked_serial, 0, 0);
 632                         matched = true;
 633                 }
 634
 635                 /* If a later serial is being acked, then mark this slot as
 636                  * being available.
 637                  */
 638                 if (after(acked_serial, orig_serial)) {
 639                         trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_obsolete, i,
 640                                            orig_serial, acked_serial, 0, 0);
 641                         clear_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
 642                         smp_wmb();
 643                         set_bit(i, &call->rtt_avail);
 644                 }
 645         }
 646
 647         if (!matched)
 648                 trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_lost, 9, 0, acked_serial, 0, 0);
 649 }
 650
 651 /*
 652  * Process the extra information that may be appended to an ACK packet
 653  */
 654 static void rxrpc_input_ackinfo(struct rxrpc_call *call, struct sk_buff *skb,
 655                                 struct rxrpc_ackinfo *ackinfo)
 656 {
 657         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 658         struct rxrpc_peer *peer;
 659         unsigned int mtu;
 660         bool wake = false;
 661         u32 rwind = ntohl(ackinfo->rwind);
 662
 663         if (rwind > RXRPC_TX_MAX_WINDOW)
 664                 rwind = RXRPC_TX_MAX_WINDOW;
 665         if (call->tx_winsize != rwind) {
 666                 if (rwind > call->tx_winsize)
 667                         wake = true;
 668                 trace_rxrpc_rx_rwind_change(call, sp->hdr.serial, rwind, wake);
 669                 call->tx_winsize = rwind;
 670         }
 671
 672         if (call->cong_ssthresh > rwind)
 673                 call->cong_ssthresh = rwind;
 674
 675         mtu = min(ntohl(ackinfo->rxMTU), ntohl(ackinfo->maxMTU));
 676
 677         peer = call->peer;
 678         if (mtu < peer->maxdata) {
 679                 spin_lock(&peer->lock);
 680                 peer->maxdata = mtu;
 681                 peer->mtu = mtu + peer->hdrsize;
 682                 spin_unlock(&peer->lock);
 683         }
 684
 685         if (wake)
 686                 wake_up(&call->waitq);
 687 }
 688
 689 /*
 690  * Process individual soft ACKs.
 691  *
 692  * Each ACK in the array corresponds to one packet and can be either an ACK or
 693  * a NAK.  If we get find an explicitly NAK'd packet we resend immediately;
 694  * packets that lie beyond the end of the ACK list are scheduled for resend by
 695  * the timer on the basis that the peer might just not have processed them at
 696  * the time the ACK was sent.
 697  */
 698 static void rxrpc_input_soft_acks(struct rxrpc_call *call, u8 *acks,
 699                                   rxrpc_seq_t seq, int nr_acks,
 700                                   struct rxrpc_ack_summary *summary)
 701 {
 702         unsigned int i;
 703
 704         for (i = 0; i < nr_acks; i++) {
 705                 if (acks[i] == RXRPC_ACK_TYPE_ACK) {
 706                         summary->nr_acks++;
 707                         summary->nr_new_acks++;
 708                 } else {
 709                         if (!summary->saw_nacks &&
 710                             call->acks_lowest_nak != seq + i) {
 711                                 call->acks_lowest_nak = seq + i;
 712                                 summary->new_low_nack = true;
 713                         }
 714                         summary->saw_nacks = true;
 715                 }
 716         }
 717 }
 718
 719 /*
 720  * Return true if the ACK is valid - ie. it doesn't appear to have regressed
 721  * with respect to the ack state conveyed by preceding ACKs.
 722  */
 723 static bool rxrpc_is_ack_valid(struct rxrpc_call *call,
 724                                rxrpc_seq_t first_pkt, rxrpc_seq_t prev_pkt)
 725 {
 726         rxrpc_seq_t base = READ_ONCE(call->acks_first_seq);
 727
 728         if (after(first_pkt, base))
 729                 return true; /* The window advanced */
 730
 731         if (before(first_pkt, base))
 732                 return false; /* firstPacket regressed */
 733
 734         if (after_eq(prev_pkt, call->acks_prev_seq))
 735                 return true; /* previousPacket hasn't regressed. */
 736
 737         /* Some rx implementations put a serial number in previousPacket. */
 738         if (after_eq(prev_pkt, base + call->tx_winsize))
 739                 return false;
 740         return true;
 741 }
 742
 743 /*
 744  * Process an ACK packet.
 745  *
 746  * ack.firstPacket is the sequence number of the first soft-ACK'd/NAK'd packet
 747  * in the ACK array.  Anything before that is hard-ACK'd and may be discarded.
 748  *
 749  * A hard-ACK means that a packet has been processed and may be discarded; a
 750  * soft-ACK means that the packet may be discarded and retransmission
 751  * requested.  A phase is complete when all packets are hard-ACK'd.
 752  */
 753 static void rxrpc_input_ack(struct rxrpc_call *call, struct sk_buff *skb)
 754 {
 755         struct rxrpc_ack_summary summary = { 0 };
 756         struct rxrpc_ackpacket ack;
 757         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 758         struct rxrpc_ackinfo info;
 759         rxrpc_serial_t ack_serial, acked_serial;
 760         rxrpc_seq_t first_soft_ack, hard_ack, prev_pkt;
 761         int nr_acks, offset, ioffset;
 762
 763         _enter("");
 764
 765         offset = sizeof(struct rxrpc_wire_header);
 766         if (skb_copy_bits(skb, offset, &ack, sizeof(ack)) < 0)
 767                 return rxrpc_proto_abort("XAK", call, 0);
 768         offset += sizeof(ack);
 769
 770         ack_serial = sp->hdr.serial;
 771         acked_serial = ntohl(ack.serial);
 772         first_soft_ack = ntohl(ack.firstPacket);
 773         prev_pkt = ntohl(ack.previousPacket);
 774         hard_ack = first_soft_ack - 1;
 775         nr_acks = ack.nAcks;
 776         summary.ack_reason = (ack.reason < RXRPC_ACK__INVALID ?
 777                               ack.reason : RXRPC_ACK__INVALID);
 778
 779         trace_rxrpc_rx_ack(call, ack_serial, acked_serial,
 780                            first_soft_ack, prev_pkt,
 781                            summary.ack_reason, nr_acks);
 782         rxrpc_inc_stat(call->rxnet, stat_rx_acks[ack.reason]);
 783
 784         switch (ack.reason) {
 785         case RXRPC_ACK_PING_RESPONSE:
 786                 rxrpc_complete_rtt_probe(call, skb->tstamp, acked_serial, ack_serial,
 787                                          rxrpc_rtt_rx_ping_response);
 788                 break;
 789         case RXRPC_ACK_REQUESTED:
 790                 rxrpc_complete_rtt_probe(call, skb->tstamp, acked_serial, ack_serial,
 791                                          rxrpc_rtt_rx_requested_ack);
 792                 break;
 793         default:
 794                 if (acked_serial != 0)
 795                         rxrpc_complete_rtt_probe(call, skb->tstamp, acked_serial, ack_serial,
 796                                                  rxrpc_rtt_rx_cancel);
 797                 break;
 798         }
 799
 800         if (ack.reason == RXRPC_ACK_PING) {
 801                 rxrpc_send_ACK(call, RXRPC_ACK_PING_RESPONSE, ack_serial,
 802                                rxrpc_propose_ack_respond_to_ping);
 803         } else if (sp->hdr.flags & RXRPC_REQUEST_ACK) {
 804                 rxrpc_send_ACK(call, RXRPC_ACK_REQUESTED, ack_serial,
 805                                rxrpc_propose_ack_respond_to_ack);
 806         }
 807
 808         /* If we get an EXCEEDS_WINDOW ACK from the server, it probably
 809          * indicates that the client address changed due to NAT.  The server
 810          * lost the call because it switched to a different peer.
 811          */
 812         if (unlikely(ack.reason == RXRPC_ACK_EXCEEDS_WINDOW) &&
 813             first_soft_ack == 1 &&
 814             prev_pkt == 0 &&
 815             rxrpc_is_client_call(call)) {
 816                 rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
 817                                           0, -ENETRESET);
 818                 return;
 819         }
 820
 821         /* If we get an OUT_OF_SEQUENCE ACK from the server, that can also
 822          * indicate a change of address.  However, we can retransmit the call
 823          * if we still have it buffered to the beginning.
 824          */
 825         if (unlikely(ack.reason == RXRPC_ACK_OUT_OF_SEQUENCE) &&
 826             first_soft_ack == 1 &&
 827             prev_pkt == 0 &&
 828             call->acks_hard_ack == 0 &&
 829             rxrpc_is_client_call(call)) {
 830                 rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
 831                                           0, -ENETRESET);
 832                 return;
 833         }
 834
 835         /* Discard any out-of-order or duplicate ACKs (outside lock). */
 836         if (!rxrpc_is_ack_valid(call, first_soft_ack, prev_pkt)) {
 837                 trace_rxrpc_rx_discard_ack(call->debug_id, ack_serial,
 838                                            first_soft_ack, call->acks_first_seq,
 839                                            prev_pkt, call->acks_prev_seq);
 840                 return;
 841         }
 842
 843         info.rxMTU = 0;
 844         ioffset = offset + nr_acks + 3;
 845         if (skb->len >= ioffset + sizeof(info) &&
 846             skb_copy_bits(skb, ioffset, &info, sizeof(info)) < 0)
 847                 return rxrpc_proto_abort("XAI", call, 0);
 848
 849         if (nr_acks > 0)
 850                 skb_condense(skb);
 851
 852         call->acks_latest_ts = skb->tstamp;
 853         call->acks_first_seq = first_soft_ack;
 854         call->acks_prev_seq = prev_pkt;
 855
 856         switch (ack.reason) {
 857         case RXRPC_ACK_PING:
 858                 break;
 859         default:
 860                 if (after(acked_serial, call->acks_highest_serial))
 861                         call->acks_highest_serial = acked_serial;
 862                 break;
 863         }
 864
 865         /* Parse rwind and mtu sizes if provided. */
 866         if (info.rxMTU)
 867                 rxrpc_input_ackinfo(call, skb, &info);
 868
 869         if (first_soft_ack == 0)
 870                 return rxrpc_proto_abort("AK0", call, 0);
 871
 872         /* Ignore ACKs unless we are or have just been transmitting. */
 873         switch (READ_ONCE(call->state)) {
 874         case RXRPC_CALL_CLIENT_SEND_REQUEST:
 875         case RXRPC_CALL_CLIENT_AWAIT_REPLY:
 876         case RXRPC_CALL_SERVER_SEND_REPLY:
 877         case RXRPC_CALL_SERVER_AWAIT_ACK:
 878                 break;
 879         default:
 880                 return;
 881         }
 882
 883         if (before(hard_ack, call->acks_hard_ack) ||
 884             after(hard_ack, call->tx_top))
 885                 return rxrpc_proto_abort("AKW", call, 0);
 886         if (nr_acks > call->tx_top - hard_ack)
 887                 return rxrpc_proto_abort("AKN", call, 0);
 888
 889         if (after(hard_ack, call->acks_hard_ack)) {
 890                 if (rxrpc_rotate_tx_window(call, hard_ack, &summary)) {
 891                         rxrpc_end_tx_phase(call, false, "ETA");
 892                         return;
 893                 }
 894         }
 895
 896         if (nr_acks > 0) {
 897                 if (offset > (int)skb->len - nr_acks)
 898                         return rxrpc_proto_abort("XSA", call, 0);
 899                 rxrpc_input_soft_acks(call, skb->data + offset, first_soft_ack,
 900                                       nr_acks, &summary);
 901         }
 902
 903         if (test_bit(RXRPC_CALL_TX_LAST, &call->flags) &&
 904             summary.nr_acks == call->tx_top - hard_ack &&
 905             rxrpc_is_client_call(call))
 906                 rxrpc_propose_ping(call, ack_serial,
 907                                    rxrpc_propose_ack_ping_for_lost_reply);
 908
 909         rxrpc_congestion_management(call, skb, &summary, acked_serial);
 910 }
 911
 912 /*
 913  * Process an ACKALL packet.
 914  */
 915 static void rxrpc_input_ackall(struct rxrpc_call *call, struct sk_buff *skb)
 916 {
 917         struct rxrpc_ack_summary summary = { 0 };
 918
 919         if (rxrpc_rotate_tx_window(call, call->tx_top, &summary))
 920                 rxrpc_end_tx_phase(call, false, "ETL");
 921 }
 922
 923 /*
 924  * Process an ABORT packet directed at a call.
 925  */
 926 static void rxrpc_input_abort(struct rxrpc_call *call, struct sk_buff *skb)
 927 {
 928         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 929
 930         trace_rxrpc_rx_abort(call, sp->hdr.serial, skb->priority);
 931
 932         rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
 933                                   skb->priority, -ECONNABORTED);
 934 }
 935
 936 /*
 937  * Process an incoming call packet.
 938  */
 939 void rxrpc_input_call_packet(struct rxrpc_call *call, struct sk_buff *skb)
 940 {
 941         struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 942         unsigned long timo;
 943
 944         _enter("%p,%p", call, skb);
 945
 946         if (sp->hdr.serviceId != call->dest_srx.srx_service)
 947                 call->dest_srx.srx_service = sp->hdr.serviceId;
 948         if ((int)sp->hdr.serial - (int)call->rx_serial > 0)
 949                 call->rx_serial = sp->hdr.serial;
 950         if (!test_bit(RXRPC_CALL_RX_HEARD, &call->flags))
 951                 set_bit(RXRPC_CALL_RX_HEARD, &call->flags);
 952
 953         timo = READ_ONCE(call->next_rx_timo);
 954         if (timo) {
 955                 unsigned long now = jiffies, expect_rx_by;
 956
 957                 expect_rx_by = now + timo;
 958                 WRITE_ONCE(call->expect_rx_by, expect_rx_by);
 959                 rxrpc_reduce_call_timer(call, expect_rx_by, now,
 960                                         rxrpc_timer_set_for_normal);
 961         }
 962
 963         switch (sp->hdr.type) {
 964         case RXRPC_PACKET_TYPE_DATA:
 965                 rxrpc_input_data(call, skb);
 966                 break;
 967
 968         case RXRPC_PACKET_TYPE_ACK:
 969                 rxrpc_input_ack(call, skb);
 970                 break;
 971
 972         case RXRPC_PACKET_TYPE_BUSY:
 973                 /* Just ignore BUSY packets from the server; the retry and
 974                  * lifespan timers will take care of business.  BUSY packets
 975                  * from the client don't make sense.
 976                  */
 977                 break;
 978
 979         case RXRPC_PACKET_TYPE_ABORT:
 980                 rxrpc_input_abort(call, skb);
 981                 break;
 982
 983         case RXRPC_PACKET_TYPE_ACKALL:
 984                 rxrpc_input_ackall(call, skb);
 985                 break;
 986
 987         default:
 988                 break;
 989         }
 990 }
 991
 992 /*
 993  * Handle a new service call on a channel implicitly completing the preceding
 994  * call on that channel.  This does not apply to client conns.
 995  *
 996  * TODO: If callNumber > call_id + 1, renegotiate security.
 997  */
 998 void rxrpc_implicit_end_call(struct rxrpc_call *call, struct sk_buff *skb)
 999 {
1000         struct rxrpc_connection *conn = call->conn;
1001
1002         switch (READ_ONCE(call->state)) {
1003         case RXRPC_CALL_SERVER_AWAIT_ACK:
1004                 rxrpc_call_completed(call);
1005                 fallthrough;
1006         case RXRPC_CALL_COMPLETE:
1007                 break;
1008         default:
1009                 rxrpc_abort_call("IMP", call, 0, RX_CALL_DEAD, -ESHUTDOWN);
1010                 trace_rxrpc_improper_term(call);
1011                 break;
1012         }
1013
1014         rxrpc_input_call_event(call, skb);
1015
1016         spin_lock(&conn->bundle->channel_lock);
1017         __rxrpc_disconnect_call(conn, call);
1018         spin_unlock(&conn->bundle->channel_lock);
1019 }