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 static void rxrpc_propose_ping(struct rxrpc_call *call, 24 bool immediate, bool background) 25 { 26 if (immediate) { 27 if (background && 28 !test_and_set_bit(RXRPC_CALL_EV_PING, &call->events)) 29 rxrpc_queue_call(call); 30 } else { 31 unsigned long now = jiffies; 32 unsigned long ping_at = now + rxrpc_idle_ack_delay; 33 34 if (time_before(ping_at, call->ping_at)) { 35 WRITE_ONCE(call->ping_at, ping_at); 36 rxrpc_reduce_call_timer(call, ping_at, now, 37 rxrpc_timer_set_for_ping); 38 } 39 } 40 } 41 42 /* 43 * propose an ACK be sent 44 */ 45 static void __rxrpc_propose_ACK(struct rxrpc_call *call, u8 ack_reason, 46 u16 skew, u32 serial, bool immediate, 47 bool background, 48 enum rxrpc_propose_ack_trace why) 49 { 50 enum rxrpc_propose_ack_outcome outcome = rxrpc_propose_ack_use; 51 unsigned long expiry = rxrpc_soft_ack_delay; 52 s8 prior = rxrpc_ack_priority[ack_reason]; 53 54 /* Pings are handled specially because we don't want to accidentally 55 * lose a ping response by subsuming it into a ping. 56 */ 57 if (ack_reason == RXRPC_ACK_PING) { 58 rxrpc_propose_ping(call, immediate, background); 59 goto trace; 60 } 61 62 /* Update DELAY, IDLE, REQUESTED and PING_RESPONSE ACK serial 63 * numbers, but we don't alter the timeout. 64 */ 65 _debug("prior %u %u vs %u %u", 66 ack_reason, prior, 67 call->ackr_reason, rxrpc_ack_priority[call->ackr_reason]); 68 if (ack_reason == call->ackr_reason) { 69 if (RXRPC_ACK_UPDATEABLE & (1 << ack_reason)) { 70 outcome = rxrpc_propose_ack_update; 71 call->ackr_serial = serial; 72 call->ackr_skew = skew; 73 } 74 if (!immediate) 75 goto trace; 76 } else if (prior > rxrpc_ack_priority[call->ackr_reason]) { 77 call->ackr_reason = ack_reason; 78 call->ackr_serial = serial; 79 call->ackr_skew = skew; 80 } else { 81 outcome = rxrpc_propose_ack_subsume; 82 } 83 84 switch (ack_reason) { 85 case RXRPC_ACK_REQUESTED: 86 if (rxrpc_requested_ack_delay < expiry) 87 expiry = rxrpc_requested_ack_delay; 88 if (serial == 1) 89 immediate = false; 90 break; 91 92 case RXRPC_ACK_DELAY: 93 if (rxrpc_soft_ack_delay < expiry) 94 expiry = rxrpc_soft_ack_delay; 95 break; 96 97 case RXRPC_ACK_IDLE: 98 if (rxrpc_idle_ack_delay < expiry) 99 expiry = rxrpc_idle_ack_delay; 100 break; 101 102 default: 103 immediate = true; 104 break; 105 } 106 107 if (test_bit(RXRPC_CALL_EV_ACK, &call->events)) { 108 _debug("already scheduled"); 109 } else if (immediate || expiry == 0) { 110 _debug("immediate ACK %lx", call->events); 111 if (!test_and_set_bit(RXRPC_CALL_EV_ACK, &call->events) && 112 background) 113 rxrpc_queue_call(call); 114 } else { 115 unsigned long now = jiffies, ack_at; 116 117 if (call->peer->rtt_usage > 0) 118 ack_at = nsecs_to_jiffies(call->peer->rtt); 119 else 120 ack_at = expiry; 121 122 ack_at += READ_ONCE(call->tx_backoff); 123 ack_at += now; 124 if (time_before(ack_at, call->ack_at)) { 125 WRITE_ONCE(call->ack_at, ack_at); 126 rxrpc_reduce_call_timer(call, ack_at, now, 127 rxrpc_timer_set_for_ack); 128 } 129 } 130 131 trace: 132 trace_rxrpc_propose_ack(call, why, ack_reason, serial, immediate, 133 background, outcome); 134 } 135 136 /* 137 * propose an ACK be sent, locking the call structure 138 */ 139 void rxrpc_propose_ACK(struct rxrpc_call *call, u8 ack_reason, 140 u16 skew, u32 serial, bool immediate, bool background, 141 enum rxrpc_propose_ack_trace why) 142 { 143 spin_lock_bh(&call->lock); 144 __rxrpc_propose_ACK(call, ack_reason, skew, serial, 145 immediate, background, why); 146 spin_unlock_bh(&call->lock); 147 } 148 149 /* 150 * Handle congestion being detected by the retransmit timeout. 151 */ 152 static void rxrpc_congestion_timeout(struct rxrpc_call *call) 153 { 154 set_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags); 155 } 156 157 /* 158 * Perform retransmission of NAK'd and unack'd packets. 159 */ 160 static void rxrpc_resend(struct rxrpc_call *call, unsigned long now_j) 161 { 162 struct sk_buff *skb; 163 unsigned long resend_at; 164 rxrpc_seq_t cursor, seq, top; 165 ktime_t now, max_age, oldest, ack_ts, timeout, min_timeo; 166 int ix; 167 u8 annotation, anno_type, retrans = 0, unacked = 0; 168 169 _enter("{%d,%d}", call->tx_hard_ack, call->tx_top); 170 171 if (call->peer->rtt_usage > 1) 172 timeout = ns_to_ktime(call->peer->rtt * 3 / 2); 173 else 174 timeout = ms_to_ktime(rxrpc_resend_timeout); 175 min_timeo = ns_to_ktime((1000000000 / HZ) * 4); 176 if (ktime_before(timeout, min_timeo)) 177 timeout = min_timeo; 178 179 now = ktime_get_real(); 180 max_age = ktime_sub(now, timeout); 181 182 spin_lock_bh(&call->lock); 183 184 cursor = call->tx_hard_ack; 185 top = call->tx_top; 186 ASSERT(before_eq(cursor, top)); 187 if (cursor == top) 188 goto out_unlock; 189 190 /* Scan the packet list without dropping the lock and decide which of 191 * the packets in the Tx buffer we're going to resend and what the new 192 * resend timeout will be. 193 */ 194 trace_rxrpc_resend(call, (cursor + 1) & RXRPC_RXTX_BUFF_MASK); 195 oldest = now; 196 for (seq = cursor + 1; before_eq(seq, top); seq++) { 197 ix = seq & RXRPC_RXTX_BUFF_MASK; 198 annotation = call->rxtx_annotations[ix]; 199 anno_type = annotation & RXRPC_TX_ANNO_MASK; 200 annotation &= ~RXRPC_TX_ANNO_MASK; 201 if (anno_type == RXRPC_TX_ANNO_ACK) 202 continue; 203 204 skb = call->rxtx_buffer[ix]; 205 rxrpc_see_skb(skb, rxrpc_skb_tx_seen); 206 207 if (anno_type == RXRPC_TX_ANNO_UNACK) { 208 if (ktime_after(skb->tstamp, max_age)) { 209 if (ktime_before(skb->tstamp, oldest)) 210 oldest = skb->tstamp; 211 continue; 212 } 213 if (!(annotation & RXRPC_TX_ANNO_RESENT)) 214 unacked++; 215 } 216 217 /* Okay, we need to retransmit a packet. */ 218 call->rxtx_annotations[ix] = RXRPC_TX_ANNO_RETRANS | annotation; 219 retrans++; 220 trace_rxrpc_retransmit(call, seq, annotation | anno_type, 221 ktime_to_ns(ktime_sub(skb->tstamp, max_age))); 222 } 223 224 resend_at = nsecs_to_jiffies(ktime_to_ns(ktime_sub(now, oldest))); 225 resend_at += jiffies + rxrpc_resend_timeout; 226 WRITE_ONCE(call->resend_at, resend_at); 227 228 if (unacked) 229 rxrpc_congestion_timeout(call); 230 231 /* If there was nothing that needed retransmission then it's likely 232 * that an ACK got lost somewhere. Send a ping to find out instead of 233 * retransmitting data. 234 */ 235 if (!retrans) { 236 rxrpc_reduce_call_timer(call, resend_at, now_j, 237 rxrpc_timer_set_for_resend); 238 spin_unlock_bh(&call->lock); 239 ack_ts = ktime_sub(now, call->acks_latest_ts); 240 if (ktime_to_ns(ack_ts) < call->peer->rtt) 241 goto out; 242 rxrpc_propose_ACK(call, RXRPC_ACK_PING, 0, 0, true, false, 243 rxrpc_propose_ack_ping_for_lost_ack); 244 rxrpc_send_ack_packet(call, true, NULL); 245 goto out; 246 } 247 248 /* Now go through the Tx window and perform the retransmissions. We 249 * have to drop the lock for each send. If an ACK comes in whilst the 250 * lock is dropped, it may clear some of the retransmission markers for 251 * packets that it soft-ACKs. 252 */ 253 for (seq = cursor + 1; before_eq(seq, top); seq++) { 254 ix = seq & RXRPC_RXTX_BUFF_MASK; 255 annotation = call->rxtx_annotations[ix]; 256 anno_type = annotation & RXRPC_TX_ANNO_MASK; 257 if (anno_type != RXRPC_TX_ANNO_RETRANS) 258 continue; 259 260 skb = call->rxtx_buffer[ix]; 261 rxrpc_get_skb(skb, rxrpc_skb_tx_got); 262 spin_unlock_bh(&call->lock); 263 264 if (rxrpc_send_data_packet(call, skb, true) < 0) { 265 rxrpc_free_skb(skb, rxrpc_skb_tx_freed); 266 return; 267 } 268 269 if (rxrpc_is_client_call(call)) 270 rxrpc_expose_client_call(call); 271 272 rxrpc_free_skb(skb, rxrpc_skb_tx_freed); 273 spin_lock_bh(&call->lock); 274 275 /* We need to clear the retransmit state, but there are two 276 * things we need to be aware of: A new ACK/NAK might have been 277 * received and the packet might have been hard-ACK'd (in which 278 * case it will no longer be in the buffer). 279 */ 280 if (after(seq, call->tx_hard_ack)) { 281 annotation = call->rxtx_annotations[ix]; 282 anno_type = annotation & RXRPC_TX_ANNO_MASK; 283 if (anno_type == RXRPC_TX_ANNO_RETRANS || 284 anno_type == RXRPC_TX_ANNO_NAK) { 285 annotation &= ~RXRPC_TX_ANNO_MASK; 286 annotation |= RXRPC_TX_ANNO_UNACK; 287 } 288 annotation |= RXRPC_TX_ANNO_RESENT; 289 call->rxtx_annotations[ix] = annotation; 290 } 291 292 if (after(call->tx_hard_ack, seq)) 293 seq = call->tx_hard_ack; 294 } 295 296 out_unlock: 297 spin_unlock_bh(&call->lock); 298 out: 299 _leave(""); 300 } 301 302 /* 303 * Handle retransmission and deferred ACK/abort generation. 304 */ 305 void rxrpc_process_call(struct work_struct *work) 306 { 307 struct rxrpc_call *call = 308 container_of(work, struct rxrpc_call, processor); 309 rxrpc_serial_t *send_ack; 310 unsigned long now, next, t; 311 unsigned int iterations = 0; 312 313 rxrpc_see_call(call); 314 315 //printk("\n--------------------\n"); 316 _enter("{%d,%s,%lx}", 317 call->debug_id, rxrpc_call_states[call->state], call->events); 318 319 recheck_state: 320 /* Limit the number of times we do this before returning to the manager */ 321 iterations++; 322 if (iterations > 5) 323 goto requeue; 324 325 if (test_and_clear_bit(RXRPC_CALL_EV_ABORT, &call->events)) { 326 rxrpc_send_abort_packet(call); 327 goto recheck_state; 328 } 329 330 if (call->state == RXRPC_CALL_COMPLETE) { 331 del_timer_sync(&call->timer); 332 rxrpc_notify_socket(call); 333 goto out_put; 334 } 335 336 /* Work out if any timeouts tripped */ 337 now = jiffies; 338 t = READ_ONCE(call->expect_rx_by); 339 if (time_after_eq(now, t)) { 340 trace_rxrpc_timer(call, rxrpc_timer_exp_normal, now); 341 set_bit(RXRPC_CALL_EV_EXPIRED, &call->events); 342 } 343 344 t = READ_ONCE(call->expect_req_by); 345 if (call->state == RXRPC_CALL_SERVER_RECV_REQUEST && 346 time_after_eq(now, t)) { 347 trace_rxrpc_timer(call, rxrpc_timer_exp_idle, now); 348 set_bit(RXRPC_CALL_EV_EXPIRED, &call->events); 349 } 350 351 t = READ_ONCE(call->expect_term_by); 352 if (time_after_eq(now, t)) { 353 trace_rxrpc_timer(call, rxrpc_timer_exp_hard, now); 354 set_bit(RXRPC_CALL_EV_EXPIRED, &call->events); 355 } 356 357 t = READ_ONCE(call->ack_at); 358 if (time_after_eq(now, t)) { 359 trace_rxrpc_timer(call, rxrpc_timer_exp_ack, now); 360 cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET); 361 set_bit(RXRPC_CALL_EV_ACK, &call->events); 362 } 363 364 t = READ_ONCE(call->ack_lost_at); 365 if (time_after_eq(now, t)) { 366 trace_rxrpc_timer(call, rxrpc_timer_exp_lost_ack, now); 367 cmpxchg(&call->ack_lost_at, t, now + MAX_JIFFY_OFFSET); 368 set_bit(RXRPC_CALL_EV_ACK_LOST, &call->events); 369 } 370 371 t = READ_ONCE(call->keepalive_at); 372 if (time_after_eq(now, t)) { 373 trace_rxrpc_timer(call, rxrpc_timer_exp_keepalive, now); 374 cmpxchg(&call->keepalive_at, t, now + MAX_JIFFY_OFFSET); 375 rxrpc_propose_ACK(call, RXRPC_ACK_PING, 0, 0, true, true, 376 rxrpc_propose_ack_ping_for_keepalive); 377 set_bit(RXRPC_CALL_EV_PING, &call->events); 378 } 379 380 t = READ_ONCE(call->ping_at); 381 if (time_after_eq(now, t)) { 382 trace_rxrpc_timer(call, rxrpc_timer_exp_ping, now); 383 cmpxchg(&call->ping_at, t, now + MAX_JIFFY_OFFSET); 384 set_bit(RXRPC_CALL_EV_PING, &call->events); 385 } 386 387 t = READ_ONCE(call->resend_at); 388 if (time_after_eq(now, t)) { 389 trace_rxrpc_timer(call, rxrpc_timer_exp_resend, now); 390 cmpxchg(&call->resend_at, t, now + MAX_JIFFY_OFFSET); 391 set_bit(RXRPC_CALL_EV_RESEND, &call->events); 392 } 393 394 /* Process events */ 395 if (test_and_clear_bit(RXRPC_CALL_EV_EXPIRED, &call->events)) { 396 if (test_bit(RXRPC_CALL_RX_HEARD, &call->flags) && 397 (int)call->conn->hi_serial - (int)call->rx_serial > 0) { 398 trace_rxrpc_call_reset(call); 399 rxrpc_abort_call("EXP", call, 0, RX_USER_ABORT, -ECONNRESET); 400 } else { 401 rxrpc_abort_call("EXP", call, 0, RX_USER_ABORT, -ETIME); 402 } 403 set_bit(RXRPC_CALL_EV_ABORT, &call->events); 404 goto recheck_state; 405 } 406 407 send_ack = NULL; 408 if (test_and_clear_bit(RXRPC_CALL_EV_ACK_LOST, &call->events)) { 409 call->acks_lost_top = call->tx_top; 410 rxrpc_propose_ACK(call, RXRPC_ACK_PING, 0, 0, true, false, 411 rxrpc_propose_ack_ping_for_lost_ack); 412 send_ack = &call->acks_lost_ping; 413 } 414 415 if (test_and_clear_bit(RXRPC_CALL_EV_ACK, &call->events) || 416 send_ack) { 417 if (call->ackr_reason) { 418 rxrpc_send_ack_packet(call, false, send_ack); 419 goto recheck_state; 420 } 421 } 422 423 if (test_and_clear_bit(RXRPC_CALL_EV_PING, &call->events)) { 424 rxrpc_send_ack_packet(call, true, NULL); 425 goto recheck_state; 426 } 427 428 if (test_and_clear_bit(RXRPC_CALL_EV_RESEND, &call->events)) { 429 rxrpc_resend(call, now); 430 goto recheck_state; 431 } 432 433 /* Make sure the timer is restarted */ 434 next = call->expect_rx_by; 435 436 #define set(T) { t = READ_ONCE(T); if (time_before(t, next)) next = t; } 437 438 set(call->expect_req_by); 439 set(call->expect_term_by); 440 set(call->ack_at); 441 set(call->ack_lost_at); 442 set(call->resend_at); 443 set(call->keepalive_at); 444 set(call->ping_at); 445 446 now = jiffies; 447 if (time_after_eq(now, next)) 448 goto recheck_state; 449 450 rxrpc_reduce_call_timer(call, next, now, rxrpc_timer_restart); 451 452 /* other events may have been raised since we started checking */ 453 if (call->events && call->state < RXRPC_CALL_COMPLETE) 454 goto requeue; 455 456 out_put: 457 rxrpc_put_call(call, rxrpc_call_put); 458 out: 459 _leave(""); 460 return; 461 462 requeue: 463 __rxrpc_queue_call(call); 464 goto out; 465 } 466