net/rxrpc/call_event.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* Management of Tx window, Tx resend, ACKs and out-of-sequence reception
   3  *
   4  * Copyright (C) 2007 Red Hat, Inc. All Rights Reserved.
   5  * Written by David Howells (dhowells@redhat.com)
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/module.h>
  11 #include <linux/circ_buf.h>
  12 #include <linux/net.h>
  13 #include <linux/skbuff.h>
  14 #include <linux/slab.h>
  15 #include <linux/udp.h>
  16 #include <net/sock.h>
  17 #include <net/af_rxrpc.h>
  18 #include "ar-internal.h"
  19
  20 /*
  21  * Propose a PING ACK be sent.
  22  */
  23 void rxrpc_propose_ping(struct rxrpc_call *call, u32 serial,
  24                         enum rxrpc_propose_ack_trace why)
  25 {
  26         unsigned long now = jiffies;
  27         unsigned long ping_at = now + rxrpc_idle_ack_delay;
  28
  29         if (time_before(ping_at, call->ping_at)) {
  30                 WRITE_ONCE(call->ping_at, ping_at);
  31                 rxrpc_reduce_call_timer(call, ping_at, now,
  32                                         rxrpc_timer_set_for_ping);
  33                 trace_rxrpc_propose_ack(call, why, RXRPC_ACK_PING, serial);
  34         }
  35 }
  36
  37 /*
  38  * Propose a DELAY ACK be sent in the future.
  39  */
  40 void rxrpc_propose_delay_ACK(struct rxrpc_call *call, rxrpc_serial_t serial,
  41                              enum rxrpc_propose_ack_trace why)
  42 {
  43         unsigned long expiry = rxrpc_soft_ack_delay;
  44         unsigned long now = jiffies, ack_at;
  45
  46         call->ackr_serial = serial;
  47
  48         if (rxrpc_soft_ack_delay < expiry)
  49                 expiry = rxrpc_soft_ack_delay;
  50         if (call->peer->srtt_us != 0)
  51                 ack_at = usecs_to_jiffies(call->peer->srtt_us >> 3);
  52         else
  53                 ack_at = expiry;
  54
  55         ack_at += READ_ONCE(call->tx_backoff);
  56         ack_at += now;
  57         if (time_before(ack_at, call->delay_ack_at)) {
  58                 WRITE_ONCE(call->delay_ack_at, ack_at);
  59                 rxrpc_reduce_call_timer(call, ack_at, now,
  60                                         rxrpc_timer_set_for_ack);
  61         }
  62
  63         trace_rxrpc_propose_ack(call, why, RXRPC_ACK_DELAY, serial);
  64 }
  65
  66 /*
  67  * Queue an ACK for immediate transmission.
  68  */
  69 void rxrpc_send_ACK(struct rxrpc_call *call, u8 ack_reason,
  70                     rxrpc_serial_t serial, enum rxrpc_propose_ack_trace why)
  71 {
  72         struct rxrpc_txbuf *txb;
  73
  74         if (test_bit(RXRPC_CALL_DISCONNECTED, &call->flags))
  75                 return;
  76
  77         rxrpc_inc_stat(call->rxnet, stat_tx_acks[ack_reason]);
  78
  79         txb = rxrpc_alloc_txbuf(call, RXRPC_PACKET_TYPE_ACK,
  80                                 rcu_read_lock_held() ? GFP_ATOMIC | __GFP_NOWARN : GFP_NOFS);
  81         if (!txb) {
  82                 kleave(" = -ENOMEM");
  83                 return;
  84         }
  85
  86         txb->ack_why            = why;
  87         txb->wire.seq           = 0;
  88         txb->wire.type          = RXRPC_PACKET_TYPE_ACK;
  89         txb->wire.flags         |= RXRPC_SLOW_START_OK;
  90         txb->ack.bufferSpace    = 0;
  91         txb->ack.maxSkew        = 0;
  92         txb->ack.firstPacket    = 0;
  93         txb->ack.previousPacket = 0;
  94         txb->ack.serial         = htonl(serial);
  95         txb->ack.reason         = ack_reason;
  96         txb->ack.nAcks          = 0;
  97
  98         trace_rxrpc_send_ack(call, why, ack_reason, serial);
  99         rxrpc_send_ack_packet(call, txb);
 100         rxrpc_put_txbuf(txb, rxrpc_txbuf_put_ack_tx);
 101 }
 102
 103 /*
 104  * Handle congestion being detected by the retransmit timeout.
 105  */
 106 static void rxrpc_congestion_timeout(struct rxrpc_call *call)
 107 {
 108         set_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags);
 109 }
 110
 111 /*
 112  * Perform retransmission of NAK'd and unack'd packets.
 113  */
 114 void rxrpc_resend(struct rxrpc_call *call, struct sk_buff *ack_skb)
 115 {
 116         struct rxrpc_ackpacket *ack = NULL;
 117         struct rxrpc_txbuf *txb;
 118         unsigned long resend_at;
 119         rxrpc_seq_t transmitted = READ_ONCE(call->tx_transmitted);
 120         ktime_t now, max_age, oldest, ack_ts;
 121         bool unacked = false;
 122         unsigned int i;
 123         LIST_HEAD(retrans_queue);
 124
 125         _enter("{%d,%d}", call->acks_hard_ack, call->tx_top);
 126
 127         now = ktime_get_real();
 128         max_age = ktime_sub_us(now, jiffies_to_usecs(call->peer->rto_j));
 129         oldest = now;
 130
 131         if (list_empty(&call->tx_buffer))
 132                 goto no_resend;
 133
 134         if (list_empty(&call->tx_buffer))
 135                 goto no_further_resend;
 136
 137         trace_rxrpc_resend(call, ack_skb);
 138         txb = list_first_entry(&call->tx_buffer, struct rxrpc_txbuf, call_link);
 139
 140         /* Scan the soft ACK table without dropping the lock and resend any
 141          * explicitly NAK'd packets.
 142          */
 143         if (ack_skb) {
 144                 ack = (void *)ack_skb->data + sizeof(struct rxrpc_wire_header);
 145
 146                 for (i = 0; i < ack->nAcks; i++) {
 147                         rxrpc_seq_t seq;
 148
 149                         if (ack->acks[i] & 1)
 150                                 continue;
 151                         seq = ntohl(ack->firstPacket) + i;
 152                         if (after(txb->seq, transmitted))
 153                                 break;
 154                         if (after(txb->seq, seq))
 155                                 continue; /* A new hard ACK probably came in */
 156                         list_for_each_entry_from(txb, &call->tx_buffer, call_link) {
 157                                 if (txb->seq == seq)
 158                                         goto found_txb;
 159                         }
 160                         goto no_further_resend;
 161
 162                 found_txb:
 163                         if (after(ntohl(txb->wire.serial), call->acks_highest_serial))
 164                                 continue; /* Ack point not yet reached */
 165
 166                         rxrpc_see_txbuf(txb, rxrpc_txbuf_see_unacked);
 167
 168                         if (list_empty(&txb->tx_link)) {
 169                                 list_add_tail(&txb->tx_link, &retrans_queue);
 170                                 set_bit(RXRPC_TXBUF_RESENT, &txb->flags);
 171                         }
 172
 173                         trace_rxrpc_retransmit(call, txb->seq,
 174                                                ktime_to_ns(ktime_sub(txb->last_sent,
 175                                                                      max_age)));
 176
 177                         if (list_is_last(&txb->call_link, &call->tx_buffer))
 178                                 goto no_further_resend;
 179                         txb = list_next_entry(txb, call_link);
 180                 }
 181         }
 182
 183         /* Fast-forward through the Tx queue to the point the peer says it has
 184          * seen.  Anything between the soft-ACK table and that point will get
 185          * ACK'd or NACK'd in due course, so don't worry about it here; here we
 186          * need to consider retransmitting anything beyond that point.
 187          *
 188          * Note that ACK for a packet can beat the update of tx_transmitted.
 189          */
 190         if (after_eq(READ_ONCE(call->acks_prev_seq), READ_ONCE(call->tx_transmitted)))
 191                 goto no_further_resend;
 192
 193         list_for_each_entry_from(txb, &call->tx_buffer, call_link) {
 194                 if (before_eq(txb->seq, READ_ONCE(call->acks_prev_seq)))
 195                         continue;
 196                 if (after(txb->seq, READ_ONCE(call->tx_transmitted)))
 197                         break; /* Not transmitted yet */
 198
 199                 if (ack && ack->reason == RXRPC_ACK_PING_RESPONSE &&
 200                     before(ntohl(txb->wire.serial), ntohl(ack->serial)))
 201                         goto do_resend; /* Wasn't accounted for by a more recent ping. */
 202
 203                 if (ktime_after(txb->last_sent, max_age)) {
 204                         if (ktime_before(txb->last_sent, oldest))
 205                                 oldest = txb->last_sent;
 206                         continue;
 207                 }
 208
 209         do_resend:
 210                 unacked = true;
 211                 if (list_empty(&txb->tx_link)) {
 212                         list_add_tail(&txb->tx_link, &retrans_queue);
 213                         set_bit(RXRPC_TXBUF_RESENT, &txb->flags);
 214                         rxrpc_inc_stat(call->rxnet, stat_tx_data_retrans);
 215                 }
 216         }
 217
 218 no_further_resend:
 219 no_resend:
 220         resend_at = nsecs_to_jiffies(ktime_to_ns(ktime_sub(now, oldest)));
 221         resend_at += jiffies + rxrpc_get_rto_backoff(call->peer,
 222                                                      !list_empty(&retrans_queue));
 223         WRITE_ONCE(call->resend_at, resend_at);
 224
 225         if (unacked)
 226                 rxrpc_congestion_timeout(call);
 227
 228         /* If there was nothing that needed retransmission then it's likely
 229          * that an ACK got lost somewhere.  Send a ping to find out instead of
 230          * retransmitting data.
 231          */
 232         if (list_empty(&retrans_queue)) {
 233                 rxrpc_reduce_call_timer(call, resend_at, jiffies,
 234                                         rxrpc_timer_set_for_resend);
 235                 ack_ts = ktime_sub(now, call->acks_latest_ts);
 236                 if (ktime_to_us(ack_ts) < (call->peer->srtt_us >> 3))
 237                         goto out;
 238                 rxrpc_send_ACK(call, RXRPC_ACK_PING, 0,
 239                                rxrpc_propose_ack_ping_for_lost_ack);
 240                 goto out;
 241         }
 242
 243         /* Retransmit the queue */
 244         while ((txb = list_first_entry_or_null(&retrans_queue,
 245                                                struct rxrpc_txbuf, tx_link))) {
 246                 list_del_init(&txb->tx_link);
 247                 rxrpc_transmit_one(call, txb);
 248         }
 249
 250 out:
 251         _leave("");
 252 }
 253
 254 /*
 255  * Start transmitting the reply to a service.  This cancels the need to ACK the
 256  * request if we haven't yet done so.
 257  */
 258 static void rxrpc_begin_service_reply(struct rxrpc_call *call)
 259 {
 260         unsigned long now;
 261
 262         write_lock(&call->state_lock);
 263
 264         if (call->state == RXRPC_CALL_SERVER_ACK_REQUEST) {
 265                 now = jiffies;
 266                 call->state = RXRPC_CALL_SERVER_SEND_REPLY;
 267                 WRITE_ONCE(call->delay_ack_at, now + MAX_JIFFY_OFFSET);
 268                 if (call->ackr_reason == RXRPC_ACK_DELAY)
 269                         call->ackr_reason = 0;
 270                 trace_rxrpc_timer(call, rxrpc_timer_init_for_send_reply, now);
 271         }
 272
 273         write_unlock(&call->state_lock);
 274 }
 275
 276 /*
 277  * Close the transmission phase.  After this point there is no more data to be
 278  * transmitted in the call.
 279  */
 280 static void rxrpc_close_tx_phase(struct rxrpc_call *call)
 281 {
 282         _debug("________awaiting reply/ACK__________");
 283
 284         write_lock(&call->state_lock);
 285         switch (call->state) {
 286         case RXRPC_CALL_CLIENT_SEND_REQUEST:
 287                 call->state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
 288                 break;
 289         case RXRPC_CALL_SERVER_SEND_REPLY:
 290                 call->state = RXRPC_CALL_SERVER_AWAIT_ACK;
 291                 break;
 292         default:
 293                 break;
 294         }
 295         write_unlock(&call->state_lock);
 296 }
 297
 298 static bool rxrpc_tx_window_has_space(struct rxrpc_call *call)
 299 {
 300         unsigned int winsize = min_t(unsigned int, call->tx_winsize,
 301                                      call->cong_cwnd + call->cong_extra);
 302         rxrpc_seq_t window = call->acks_hard_ack, wtop = window + winsize;
 303         rxrpc_seq_t tx_top = call->tx_top;
 304         int space;
 305
 306         space = wtop - tx_top;
 307         return space > 0;
 308 }
 309
 310 /*
 311  * Decant some if the sendmsg prepared queue into the transmission buffer.
 312  */
 313 static void rxrpc_decant_prepared_tx(struct rxrpc_call *call)
 314 {
 315         struct rxrpc_txbuf *txb;
 316
 317         if (!test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 318                 if (list_empty(&call->tx_sendmsg))
 319                         return;
 320                 rxrpc_expose_client_call(call);
 321         }
 322
 323         while ((txb = list_first_entry_or_null(&call->tx_sendmsg,
 324                                                struct rxrpc_txbuf, call_link))) {
 325                 spin_lock(&call->tx_lock);
 326                 list_del(&txb->call_link);
 327                 spin_unlock(&call->tx_lock);
 328
 329                 call->tx_top = txb->seq;
 330                 list_add_tail(&txb->call_link, &call->tx_buffer);
 331
 332                 if (txb->wire.flags & RXRPC_LAST_PACKET)
 333                         rxrpc_close_tx_phase(call);
 334
 335                 rxrpc_transmit_one(call, txb);
 336
 337                 if (!rxrpc_tx_window_has_space(call))
 338                         break;
 339         }
 340 }
 341
 342 static void rxrpc_transmit_some_data(struct rxrpc_call *call)
 343 {
 344         switch (call->state) {
 345         case RXRPC_CALL_SERVER_ACK_REQUEST:
 346                 if (list_empty(&call->tx_sendmsg))
 347                         return;
 348                 rxrpc_begin_service_reply(call);
 349                 fallthrough;
 350
 351         case RXRPC_CALL_SERVER_SEND_REPLY:
 352         case RXRPC_CALL_CLIENT_SEND_REQUEST:
 353                 if (!rxrpc_tx_window_has_space(call))
 354                         return;
 355                 if (list_empty(&call->tx_sendmsg)) {
 356                         rxrpc_inc_stat(call->rxnet, stat_tx_data_underflow);
 357                         return;
 358                 }
 359                 rxrpc_decant_prepared_tx(call);
 360                 break;
 361         default:
 362                 return;
 363         }
 364 }
 365
 366 /*
 367  * Ping the other end to fill our RTT cache and to retrieve the rwind
 368  * and MTU parameters.
 369  */
 370 static void rxrpc_send_initial_ping(struct rxrpc_call *call)
 371 {
 372         if (call->peer->rtt_count < 3 ||
 373             ktime_before(ktime_add_ms(call->peer->rtt_last_req, 1000),
 374                          ktime_get_real()))
 375                 rxrpc_send_ACK(call, RXRPC_ACK_PING, 0,
 376                                rxrpc_propose_ack_ping_for_params);
 377 }
 378
 379 /*
 380  * Handle retransmission and deferred ACK/abort generation.
 381  */
 382 bool rxrpc_input_call_event(struct rxrpc_call *call, struct sk_buff *skb)
 383 {
 384         unsigned long now, next, t;
 385         rxrpc_serial_t ackr_serial;
 386         bool resend = false, expired = false;
 387         s32 abort_code;
 388
 389         rxrpc_see_call(call, rxrpc_call_see_input);
 390
 391         //printk("\n--------------------\n");
 392         _enter("{%d,%s,%lx}",
 393                call->debug_id, rxrpc_call_states[call->state], call->events);
 394
 395         if (call->state == RXRPC_CALL_COMPLETE)
 396                 goto out;
 397
 398         /* Handle abort request locklessly, vs rxrpc_propose_abort(). */
 399         abort_code = smp_load_acquire(&call->send_abort);
 400         if (abort_code) {
 401                 rxrpc_abort_call(call, 0, call->send_abort, call->send_abort_err,
 402                                  call->send_abort_why);
 403                 goto out;
 404         }
 405
 406         if (skb && skb->mark == RXRPC_SKB_MARK_ERROR)
 407                 goto out;
 408
 409         /* If we see our async-event poke, check for timeout trippage. */
 410         now = jiffies;
 411         t = READ_ONCE(call->expect_rx_by);
 412         if (time_after_eq(now, t)) {
 413                 trace_rxrpc_timer(call, rxrpc_timer_exp_normal, now);
 414                 expired = true;
 415         }
 416
 417         t = READ_ONCE(call->expect_req_by);
 418         if (call->state == RXRPC_CALL_SERVER_RECV_REQUEST &&
 419             time_after_eq(now, t)) {
 420                 trace_rxrpc_timer(call, rxrpc_timer_exp_idle, now);
 421                 expired = true;
 422         }
 423
 424         t = READ_ONCE(call->expect_term_by);
 425         if (time_after_eq(now, t)) {
 426                 trace_rxrpc_timer(call, rxrpc_timer_exp_hard, now);
 427                 expired = true;
 428         }
 429
 430         t = READ_ONCE(call->delay_ack_at);
 431         if (time_after_eq(now, t)) {
 432                 trace_rxrpc_timer(call, rxrpc_timer_exp_ack, now);
 433                 cmpxchg(&call->delay_ack_at, t, now + MAX_JIFFY_OFFSET);
 434                 ackr_serial = xchg(&call->ackr_serial, 0);
 435                 rxrpc_send_ACK(call, RXRPC_ACK_DELAY, ackr_serial,
 436                                rxrpc_propose_ack_ping_for_lost_ack);
 437         }
 438
 439         t = READ_ONCE(call->ack_lost_at);
 440         if (time_after_eq(now, t)) {
 441                 trace_rxrpc_timer(call, rxrpc_timer_exp_lost_ack, now);
 442                 cmpxchg(&call->ack_lost_at, t, now + MAX_JIFFY_OFFSET);
 443                 set_bit(RXRPC_CALL_EV_ACK_LOST, &call->events);
 444         }
 445
 446         t = READ_ONCE(call->keepalive_at);
 447         if (time_after_eq(now, t)) {
 448                 trace_rxrpc_timer(call, rxrpc_timer_exp_keepalive, now);
 449                 cmpxchg(&call->keepalive_at, t, now + MAX_JIFFY_OFFSET);
 450                 rxrpc_send_ACK(call, RXRPC_ACK_PING, 0,
 451                                rxrpc_propose_ack_ping_for_keepalive);
 452         }
 453
 454         t = READ_ONCE(call->ping_at);
 455         if (time_after_eq(now, t)) {
 456                 trace_rxrpc_timer(call, rxrpc_timer_exp_ping, now);
 457                 cmpxchg(&call->ping_at, t, now + MAX_JIFFY_OFFSET);
 458                 rxrpc_send_ACK(call, RXRPC_ACK_PING, 0,
 459                                rxrpc_propose_ack_ping_for_keepalive);
 460         }
 461
 462         t = READ_ONCE(call->resend_at);
 463         if (time_after_eq(now, t)) {
 464                 trace_rxrpc_timer(call, rxrpc_timer_exp_resend, now);
 465                 cmpxchg(&call->resend_at, t, now + MAX_JIFFY_OFFSET);
 466                 resend = true;
 467         }
 468
 469         if (skb)
 470                 rxrpc_input_call_packet(call, skb);
 471
 472         rxrpc_transmit_some_data(call);
 473
 474         if (skb) {
 475                 struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 476
 477                 if (sp->hdr.type == RXRPC_PACKET_TYPE_ACK)
 478                         rxrpc_congestion_degrade(call);
 479         }
 480
 481         if (test_and_clear_bit(RXRPC_CALL_EV_INITIAL_PING, &call->events))
 482                 rxrpc_send_initial_ping(call);
 483
 484         /* Process events */
 485         if (expired) {
 486                 if (test_bit(RXRPC_CALL_RX_HEARD, &call->flags) &&
 487                     (int)call->conn->hi_serial - (int)call->rx_serial > 0) {
 488                         trace_rxrpc_call_reset(call);
 489                         rxrpc_abort_call(call, 0, RX_CALL_DEAD, -ECONNRESET,
 490                                          rxrpc_abort_call_reset);
 491                 } else {
 492                         rxrpc_abort_call(call, 0, RX_CALL_TIMEOUT, -ETIME,
 493                                          rxrpc_abort_call_timeout);
 494                 }
 495                 goto out;
 496         }
 497
 498         if (test_and_clear_bit(RXRPC_CALL_EV_ACK_LOST, &call->events))
 499                 rxrpc_send_ACK(call, RXRPC_ACK_PING, 0,
 500                                rxrpc_propose_ack_ping_for_lost_ack);
 501
 502         if (resend && call->state != RXRPC_CALL_CLIENT_RECV_REPLY)
 503                 rxrpc_resend(call, NULL);
 504
 505         if (test_and_clear_bit(RXRPC_CALL_RX_IS_IDLE, &call->flags))
 506                 rxrpc_send_ACK(call, RXRPC_ACK_IDLE, 0,
 507                                rxrpc_propose_ack_rx_idle);
 508
 509         if (atomic_read(&call->ackr_nr_unacked) > 2)
 510                 rxrpc_send_ACK(call, RXRPC_ACK_IDLE, 0,
 511                                rxrpc_propose_ack_input_data);
 512
 513         /* Make sure the timer is restarted */
 514         if (call->state != RXRPC_CALL_COMPLETE) {
 515                 next = call->expect_rx_by;
 516
 517 #define set(T) { t = READ_ONCE(T); if (time_before(t, next)) next = t; }
 518
 519                 set(call->expect_req_by);
 520                 set(call->expect_term_by);
 521                 set(call->delay_ack_at);
 522                 set(call->ack_lost_at);
 523                 set(call->resend_at);
 524                 set(call->keepalive_at);
 525                 set(call->ping_at);
 526
 527                 now = jiffies;
 528                 if (time_after_eq(now, next))
 529                         rxrpc_poke_call(call, rxrpc_call_poke_timer_now);
 530
 531                 rxrpc_reduce_call_timer(call, next, now, rxrpc_timer_restart);
 532         }
 533
 534 out:
 535         if (call->state == RXRPC_CALL_COMPLETE) {
 536                 del_timer_sync(&call->timer);
 537                 if (!test_bit(RXRPC_CALL_DISCONNECTED, &call->flags))
 538                         rxrpc_disconnect_call(call);
 539                 if (call->security)
 540                         call->security->free_call_crypto(call);
 541         }
 542         if (call->acks_hard_ack != call->tx_bottom)
 543                 rxrpc_shrink_call_tx_buffer(call);
 544         _leave("");
 545         return true;
 546 }