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_local *local = call->conn->local;
  73         struct rxrpc_txbuf *txb;
  74
  75         if (test_bit(RXRPC_CALL_DISCONNECTED, &call->flags))
  76                 return;
  77         if (ack_reason == RXRPC_ACK_DELAY &&
  78             test_and_set_bit(RXRPC_CALL_DELAY_ACK_PENDING, &call->flags)) {
  79                 trace_rxrpc_drop_ack(call, why, ack_reason, serial, false);
  80                 return;
  81         }
  82
  83         rxrpc_inc_stat(call->rxnet, stat_tx_acks[ack_reason]);
  84
  85         txb = rxrpc_alloc_txbuf(call, RXRPC_PACKET_TYPE_ACK,
  86                                 rcu_read_lock_held() ? GFP_ATOMIC | __GFP_NOWARN : GFP_NOFS);
  87         if (!txb) {
  88                 kleave(" = -ENOMEM");
  89                 return;
  90         }
  91
  92         txb->ack_why            = why;
  93         txb->wire.seq           = 0;
  94         txb->wire.type          = RXRPC_PACKET_TYPE_ACK;
  95         txb->wire.flags         |= RXRPC_SLOW_START_OK;
  96         txb->ack.bufferSpace    = 0;
  97         txb->ack.maxSkew        = 0;
  98         txb->ack.firstPacket    = 0;
  99         txb->ack.previousPacket = 0;
 100         txb->ack.serial         = htonl(serial);
 101         txb->ack.reason         = ack_reason;
 102         txb->ack.nAcks          = 0;
 103
 104         if (!rxrpc_try_get_call(call, rxrpc_call_get_send_ack)) {
 105                 rxrpc_put_txbuf(txb, rxrpc_txbuf_put_nomem);
 106                 return;
 107         }
 108
 109         spin_lock_bh(&local->ack_tx_lock);
 110         list_add_tail(&txb->tx_link, &local->ack_tx_queue);
 111         spin_unlock_bh(&local->ack_tx_lock);
 112         trace_rxrpc_send_ack(call, why, ack_reason, serial);
 113
 114         if (!rcu_read_lock_held()) {
 115                 rxrpc_transmit_ack_packets(call->peer->local);
 116         } else {
 117                 rxrpc_get_local(local, rxrpc_local_get_queue);
 118                 rxrpc_queue_local(local);
 119         }
 120 }
 121
 122 /*
 123  * Handle congestion being detected by the retransmit timeout.
 124  */
 125 static void rxrpc_congestion_timeout(struct rxrpc_call *call)
 126 {
 127         set_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags);
 128 }
 129
 130 /*
 131  * Perform retransmission of NAK'd and unack'd packets.
 132  */
 133 static void rxrpc_resend(struct rxrpc_call *call, unsigned long now_j)
 134 {
 135         struct rxrpc_ackpacket *ack = NULL;
 136         struct rxrpc_txbuf *txb;
 137         struct sk_buff *ack_skb = NULL;
 138         unsigned long resend_at;
 139         rxrpc_seq_t transmitted = READ_ONCE(call->tx_transmitted);
 140         ktime_t now, max_age, oldest, ack_ts;
 141         bool unacked = false;
 142         unsigned int i;
 143         LIST_HEAD(retrans_queue);
 144
 145         _enter("{%d,%d}", call->acks_hard_ack, call->tx_top);
 146
 147         now = ktime_get_real();
 148         max_age = ktime_sub_us(now, jiffies_to_usecs(call->peer->rto_j));
 149         oldest = now;
 150
 151         /* See if there's an ACK saved with a soft-ACK table in it. */
 152         if (call->acks_soft_tbl) {
 153                 spin_lock_bh(&call->acks_ack_lock);
 154                 ack_skb = call->acks_soft_tbl;
 155                 if (ack_skb) {
 156                         rxrpc_get_skb(ack_skb, rxrpc_skb_get_ack);
 157                         ack = (void *)ack_skb->data + sizeof(struct rxrpc_wire_header);
 158                 }
 159                 spin_unlock_bh(&call->acks_ack_lock);
 160         }
 161
 162         if (list_empty(&call->tx_buffer))
 163                 goto no_resend;
 164
 165         spin_lock(&call->tx_lock);
 166
 167         if (list_empty(&call->tx_buffer))
 168                 goto no_further_resend;
 169
 170         trace_rxrpc_resend(call);
 171         txb = list_first_entry(&call->tx_buffer, struct rxrpc_txbuf, call_link);
 172
 173         /* Scan the soft ACK table without dropping the lock and resend any
 174          * explicitly NAK'd packets.
 175          */
 176         if (ack) {
 177                 for (i = 0; i < ack->nAcks; i++) {
 178                         rxrpc_seq_t seq;
 179
 180                         if (ack->acks[i] & 1)
 181                                 continue;
 182                         seq = ntohl(ack->firstPacket) + i;
 183                         if (after(txb->seq, transmitted))
 184                                 break;
 185                         if (after(txb->seq, seq))
 186                                 continue; /* A new hard ACK probably came in */
 187                         list_for_each_entry_from(txb, &call->tx_buffer, call_link) {
 188                                 if (txb->seq == seq)
 189                                         goto found_txb;
 190                         }
 191                         goto no_further_resend;
 192
 193                 found_txb:
 194                         if (after(ntohl(txb->wire.serial), call->acks_highest_serial))
 195                                 continue; /* Ack point not yet reached */
 196
 197                         rxrpc_see_txbuf(txb, rxrpc_txbuf_see_unacked);
 198
 199                         if (list_empty(&txb->tx_link)) {
 200                                 rxrpc_get_txbuf(txb, rxrpc_txbuf_get_retrans);
 201                                 list_add_tail(&txb->tx_link, &retrans_queue);
 202                                 set_bit(RXRPC_TXBUF_RESENT, &txb->flags);
 203                         }
 204
 205                         trace_rxrpc_retransmit(call, txb->seq,
 206                                                ktime_to_ns(ktime_sub(txb->last_sent,
 207                                                                      max_age)));
 208
 209                         if (list_is_last(&txb->call_link, &call->tx_buffer))
 210                                 goto no_further_resend;
 211                         txb = list_next_entry(txb, call_link);
 212                 }
 213         }
 214
 215         /* Fast-forward through the Tx queue to the point the peer says it has
 216          * seen.  Anything between the soft-ACK table and that point will get
 217          * ACK'd or NACK'd in due course, so don't worry about it here; here we
 218          * need to consider retransmitting anything beyond that point.
 219          *
 220          * Note that ACK for a packet can beat the update of tx_transmitted.
 221          */
 222         if (after_eq(READ_ONCE(call->acks_prev_seq), READ_ONCE(call->tx_transmitted)))
 223                 goto no_further_resend;
 224
 225         list_for_each_entry_from(txb, &call->tx_buffer, call_link) {
 226                 if (before_eq(txb->seq, READ_ONCE(call->acks_prev_seq)))
 227                         continue;
 228                 if (after(txb->seq, READ_ONCE(call->tx_transmitted)))
 229                         break; /* Not transmitted yet */
 230
 231                 if (ack && ack->reason == RXRPC_ACK_PING_RESPONSE &&
 232                     before(ntohl(txb->wire.serial), ntohl(ack->serial)))
 233                         goto do_resend; /* Wasn't accounted for by a more recent ping. */
 234
 235                 if (ktime_after(txb->last_sent, max_age)) {
 236                         if (ktime_before(txb->last_sent, oldest))
 237                                 oldest = txb->last_sent;
 238                         continue;
 239                 }
 240
 241         do_resend:
 242                 unacked = true;
 243                 if (list_empty(&txb->tx_link)) {
 244                         rxrpc_get_txbuf(txb, rxrpc_txbuf_get_retrans);
 245                         list_add_tail(&txb->tx_link, &retrans_queue);
 246                         set_bit(RXRPC_TXBUF_RESENT, &txb->flags);
 247                         rxrpc_inc_stat(call->rxnet, stat_tx_data_retrans);
 248                 }
 249         }
 250
 251 no_further_resend:
 252         spin_unlock(&call->tx_lock);
 253 no_resend:
 254         rxrpc_free_skb(ack_skb, rxrpc_skb_put_ack);
 255
 256         resend_at = nsecs_to_jiffies(ktime_to_ns(ktime_sub(now, oldest)));
 257         resend_at += jiffies + rxrpc_get_rto_backoff(call->peer,
 258                                                      !list_empty(&retrans_queue));
 259         WRITE_ONCE(call->resend_at, resend_at);
 260
 261         if (unacked)
 262                 rxrpc_congestion_timeout(call);
 263
 264         /* If there was nothing that needed retransmission then it's likely
 265          * that an ACK got lost somewhere.  Send a ping to find out instead of
 266          * retransmitting data.
 267          */
 268         if (list_empty(&retrans_queue)) {
 269                 rxrpc_reduce_call_timer(call, resend_at, now_j,
 270                                         rxrpc_timer_set_for_resend);
 271                 ack_ts = ktime_sub(now, call->acks_latest_ts);
 272                 if (ktime_to_us(ack_ts) < (call->peer->srtt_us >> 3))
 273                         goto out;
 274                 rxrpc_send_ACK(call, RXRPC_ACK_PING, 0,
 275                                rxrpc_propose_ack_ping_for_lost_ack);
 276                 goto out;
 277         }
 278
 279         while ((txb = list_first_entry_or_null(&retrans_queue,
 280                                                struct rxrpc_txbuf, tx_link))) {
 281                 list_del_init(&txb->tx_link);
 282                 rxrpc_send_data_packet(call, txb);
 283                 rxrpc_put_txbuf(txb, rxrpc_txbuf_put_trans);
 284
 285                 trace_rxrpc_retransmit(call, txb->seq,
 286                                        ktime_to_ns(ktime_sub(txb->last_sent,
 287                                                              max_age)));
 288         }
 289
 290 out:
 291         _leave("");
 292 }
 293
 294 static bool rxrpc_tx_window_has_space(struct rxrpc_call *call)
 295 {
 296         unsigned int winsize = min_t(unsigned int, call->tx_winsize,
 297                                      call->cong_cwnd + call->cong_extra);
 298         rxrpc_seq_t window = call->acks_hard_ack, wtop = window + winsize;
 299         rxrpc_seq_t tx_top = call->tx_top;
 300         int space;
 301
 302         space = wtop - tx_top;
 303         return space > 0;
 304 }
 305
 306 /*
 307  * Decant some if the sendmsg prepared queue into the transmission buffer.
 308  */
 309 static void rxrpc_decant_prepared_tx(struct rxrpc_call *call)
 310 {
 311         struct rxrpc_txbuf *txb;
 312
 313         if (rxrpc_is_client_call(call) &&
 314             !test_bit(RXRPC_CALL_EXPOSED, &call->flags))
 315                 rxrpc_expose_client_call(call);
 316
 317         while ((txb = list_first_entry_or_null(&call->tx_sendmsg,
 318                                                struct rxrpc_txbuf, call_link))) {
 319                 spin_lock(&call->tx_lock);
 320                 list_del(&txb->call_link);
 321                 spin_unlock(&call->tx_lock);
 322
 323                 call->tx_top = txb->seq;
 324                 list_add_tail(&txb->call_link, &call->tx_buffer);
 325
 326                 rxrpc_transmit_one(call, txb);
 327
 328                 // TODO: Drain the transmission buffers.  Do this somewhere better
 329                 if (after(call->acks_hard_ack, call->tx_bottom + 16))
 330                         rxrpc_shrink_call_tx_buffer(call);
 331
 332                 if (!rxrpc_tx_window_has_space(call))
 333                         break;
 334         }
 335 }
 336
 337 static void rxrpc_transmit_some_data(struct rxrpc_call *call)
 338 {
 339         switch (call->state) {
 340         case RXRPC_CALL_SERVER_ACK_REQUEST:
 341                 if (list_empty(&call->tx_sendmsg))
 342                         return;
 343                 fallthrough;
 344
 345         case RXRPC_CALL_SERVER_SEND_REPLY:
 346         case RXRPC_CALL_SERVER_AWAIT_ACK:
 347         case RXRPC_CALL_CLIENT_SEND_REQUEST:
 348         case RXRPC_CALL_CLIENT_AWAIT_REPLY:
 349                 if (!rxrpc_tx_window_has_space(call))
 350                         return;
 351                 if (list_empty(&call->tx_sendmsg))
 352                         return;
 353                 rxrpc_decant_prepared_tx(call);
 354                 break;
 355         default:
 356                 return;
 357         }
 358 }
 359
 360 /*
 361  * Handle retransmission and deferred ACK/abort generation.
 362  */
 363 void rxrpc_process_call(struct work_struct *work)
 364 {
 365         struct rxrpc_call *call =
 366                 container_of(work, struct rxrpc_call, processor);
 367         unsigned long now, next, t;
 368         unsigned int iterations = 0;
 369         rxrpc_serial_t ackr_serial;
 370
 371         rxrpc_see_call(call, rxrpc_call_see_input);
 372
 373         //printk("\n--------------------\n");
 374         _enter("{%d,%s,%lx}",
 375                call->debug_id, rxrpc_call_states[call->state], call->events);
 376
 377 recheck_state:
 378         if (call->acks_hard_ack != call->tx_bottom)
 379                 rxrpc_shrink_call_tx_buffer(call);
 380
 381         /* Limit the number of times we do this before returning to the manager */
 382         if (!rxrpc_tx_window_has_space(call) ||
 383             list_empty(&call->tx_sendmsg)) {
 384                 iterations++;
 385                 if (iterations > 5)
 386                         goto requeue;
 387         }
 388
 389         if (test_and_clear_bit(RXRPC_CALL_EV_ABORT, &call->events)) {
 390                 rxrpc_send_abort_packet(call);
 391                 goto recheck_state;
 392         }
 393
 394         if (call->state == RXRPC_CALL_COMPLETE) {
 395                 del_timer_sync(&call->timer);
 396                 goto out;
 397         }
 398
 399         /* Work out if any timeouts tripped */
 400         now = jiffies;
 401         t = READ_ONCE(call->expect_rx_by);
 402         if (time_after_eq(now, t)) {
 403                 trace_rxrpc_timer(call, rxrpc_timer_exp_normal, now);
 404                 set_bit(RXRPC_CALL_EV_EXPIRED, &call->events);
 405         }
 406
 407         t = READ_ONCE(call->expect_req_by);
 408         if (call->state == RXRPC_CALL_SERVER_RECV_REQUEST &&
 409             time_after_eq(now, t)) {
 410                 trace_rxrpc_timer(call, rxrpc_timer_exp_idle, now);
 411                 set_bit(RXRPC_CALL_EV_EXPIRED, &call->events);
 412         }
 413
 414         t = READ_ONCE(call->expect_term_by);
 415         if (time_after_eq(now, t)) {
 416                 trace_rxrpc_timer(call, rxrpc_timer_exp_hard, now);
 417                 set_bit(RXRPC_CALL_EV_EXPIRED, &call->events);
 418         }
 419
 420         t = READ_ONCE(call->delay_ack_at);
 421         if (time_after_eq(now, t)) {
 422                 trace_rxrpc_timer(call, rxrpc_timer_exp_ack, now);
 423                 cmpxchg(&call->delay_ack_at, t, now + MAX_JIFFY_OFFSET);
 424                 ackr_serial = xchg(&call->ackr_serial, 0);
 425                 rxrpc_send_ACK(call, RXRPC_ACK_DELAY, ackr_serial,
 426                                rxrpc_propose_ack_ping_for_lost_ack);
 427         }
 428
 429         t = READ_ONCE(call->ack_lost_at);
 430         if (time_after_eq(now, t)) {
 431                 trace_rxrpc_timer(call, rxrpc_timer_exp_lost_ack, now);
 432                 cmpxchg(&call->ack_lost_at, t, now + MAX_JIFFY_OFFSET);
 433                 set_bit(RXRPC_CALL_EV_ACK_LOST, &call->events);
 434         }
 435
 436         t = READ_ONCE(call->keepalive_at);
 437         if (time_after_eq(now, t)) {
 438                 trace_rxrpc_timer(call, rxrpc_timer_exp_keepalive, now);
 439                 cmpxchg(&call->keepalive_at, t, now + MAX_JIFFY_OFFSET);
 440                 rxrpc_send_ACK(call, RXRPC_ACK_PING, 0,
 441                                rxrpc_propose_ack_ping_for_keepalive);
 442         }
 443
 444         t = READ_ONCE(call->ping_at);
 445         if (time_after_eq(now, t)) {
 446                 trace_rxrpc_timer(call, rxrpc_timer_exp_ping, now);
 447                 cmpxchg(&call->ping_at, t, now + MAX_JIFFY_OFFSET);
 448                 rxrpc_send_ACK(call, RXRPC_ACK_PING, 0,
 449                                rxrpc_propose_ack_ping_for_keepalive);
 450         }
 451
 452         t = READ_ONCE(call->resend_at);
 453         if (time_after_eq(now, t)) {
 454                 trace_rxrpc_timer(call, rxrpc_timer_exp_resend, now);
 455                 cmpxchg(&call->resend_at, t, now + MAX_JIFFY_OFFSET);
 456                 set_bit(RXRPC_CALL_EV_RESEND, &call->events);
 457         }
 458
 459         rxrpc_transmit_some_data(call);
 460
 461         /* Process events */
 462         if (test_and_clear_bit(RXRPC_CALL_EV_EXPIRED, &call->events)) {
 463                 if (test_bit(RXRPC_CALL_RX_HEARD, &call->flags) &&
 464                     (int)call->conn->hi_serial - (int)call->rx_serial > 0) {
 465                         trace_rxrpc_call_reset(call);
 466                         rxrpc_abort_call("EXP", call, 0, RX_CALL_DEAD, -ECONNRESET);
 467                 } else {
 468                         rxrpc_abort_call("EXP", call, 0, RX_CALL_TIMEOUT, -ETIME);
 469                 }
 470                 set_bit(RXRPC_CALL_EV_ABORT, &call->events);
 471                 goto recheck_state;
 472         }
 473
 474         if (test_and_clear_bit(RXRPC_CALL_EV_ACK_LOST, &call->events)) {
 475                 call->acks_lost_top = call->tx_top;
 476                 rxrpc_send_ACK(call, RXRPC_ACK_PING, 0,
 477                                rxrpc_propose_ack_ping_for_lost_ack);
 478         }
 479
 480         if (test_and_clear_bit(RXRPC_CALL_EV_RESEND, &call->events) &&
 481             call->state != RXRPC_CALL_CLIENT_RECV_REPLY) {
 482                 rxrpc_resend(call, now);
 483                 goto recheck_state;
 484         }
 485
 486         /* Make sure the timer is restarted */
 487         next = call->expect_rx_by;
 488
 489 #define set(T) { t = READ_ONCE(T); if (time_before(t, next)) next = t; }
 490
 491         set(call->expect_req_by);
 492         set(call->expect_term_by);
 493         set(call->delay_ack_at);
 494         set(call->ack_lost_at);
 495         set(call->resend_at);
 496         set(call->keepalive_at);
 497         set(call->ping_at);
 498
 499         now = jiffies;
 500         if (time_after_eq(now, next))
 501                 goto recheck_state;
 502
 503         rxrpc_reduce_call_timer(call, next, now, rxrpc_timer_restart);
 504
 505         /* other events may have been raised since we started checking */
 506         if (call->events)
 507                 goto requeue;
 508
 509 out:
 510         _leave("");
 511         return;
 512
 513 requeue:
 514         if (call->state < RXRPC_CALL_COMPLETE)
 515                 rxrpc_queue_call(call, rxrpc_call_queue_requeue);
 516         goto out;
 517 }