1 /* Management of Tx window, Tx resend, ACKs and out-of-sequence reception 2 * 3 * Copyright (C) 2007 Red Hat, Inc. All Rights Reserved. 4 * Written by David Howells (dhowells@redhat.com) 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public License 8 * as published by the Free Software Foundation; either version 9 * 2 of the License, or (at your option) any later version. 10 */ 11 12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 13 14 #include <linux/module.h> 15 #include <linux/circ_buf.h> 16 #include <linux/net.h> 17 #include <linux/skbuff.h> 18 #include <linux/slab.h> 19 #include <linux/udp.h> 20 #include <net/sock.h> 21 #include <net/af_rxrpc.h> 22 #include "ar-internal.h" 23 24 /* 25 * Propose a PING ACK be sent. 26 */ 27 static void rxrpc_propose_ping(struct rxrpc_call *call, 28 bool immediate, bool background) 29 { 30 if (immediate) { 31 if (background && 32 !test_and_set_bit(RXRPC_CALL_EV_PING, &call->events)) 33 rxrpc_queue_call(call); 34 } else { 35 unsigned long now = jiffies; 36 unsigned long ping_at = now + rxrpc_idle_ack_delay; 37 38 if (time_before(ping_at, call->ping_at)) { 39 WRITE_ONCE(call->ping_at, ping_at); 40 rxrpc_reduce_call_timer(call, ping_at, now, 41 rxrpc_timer_set_for_ping); 42 } 43 } 44 } 45 46 /* 47 * propose an ACK be sent 48 */ 49 static void __rxrpc_propose_ACK(struct rxrpc_call *call, u8 ack_reason, 50 u16 skew, u32 serial, bool immediate, 51 bool background, 52 enum rxrpc_propose_ack_trace why) 53 { 54 enum rxrpc_propose_ack_outcome outcome = rxrpc_propose_ack_use; 55 unsigned long expiry = rxrpc_soft_ack_delay; 56 s8 prior = rxrpc_ack_priority[ack_reason]; 57 58 /* Pings are handled specially because we don't want to accidentally 59 * lose a ping response by subsuming it into a ping. 60 */ 61 if (ack_reason == RXRPC_ACK_PING) { 62 rxrpc_propose_ping(call, immediate, background); 63 goto trace; 64 } 65 66 /* Update DELAY, IDLE, REQUESTED and PING_RESPONSE ACK serial 67 * numbers, but we don't alter the timeout. 68 */ 69 _debug("prior %u %u vs %u %u", 70 ack_reason, prior, 71 call->ackr_reason, rxrpc_ack_priority[call->ackr_reason]); 72 if (ack_reason == call->ackr_reason) { 73 if (RXRPC_ACK_UPDATEABLE & (1 << ack_reason)) { 74 outcome = rxrpc_propose_ack_update; 75 call->ackr_serial = serial; 76 call->ackr_skew = skew; 77 } 78 if (!immediate) 79 goto trace; 80 } else if (prior > rxrpc_ack_priority[call->ackr_reason]) { 81 call->ackr_reason = ack_reason; 82 call->ackr_serial = serial; 83 call->ackr_skew = skew; 84 } else { 85 outcome = rxrpc_propose_ack_subsume; 86 } 87 88 switch (ack_reason) { 89 case RXRPC_ACK_REQUESTED: 90 if (rxrpc_requested_ack_delay < expiry) 91 expiry = rxrpc_requested_ack_delay; 92 if (serial == 1) 93 immediate = false; 94 break; 95 96 case RXRPC_ACK_DELAY: 97 if (rxrpc_soft_ack_delay < expiry) 98 expiry = rxrpc_soft_ack_delay; 99 break; 100 101 case RXRPC_ACK_IDLE: 102 if (rxrpc_idle_ack_delay < expiry) 103 expiry = rxrpc_idle_ack_delay; 104 break; 105 106 default: 107 immediate = true; 108 break; 109 } 110 111 if (test_bit(RXRPC_CALL_EV_ACK, &call->events)) { 112 _debug("already scheduled"); 113 } else if (immediate || expiry == 0) { 114 _debug("immediate ACK %lx", call->events); 115 if (!test_and_set_bit(RXRPC_CALL_EV_ACK, &call->events) && 116 background) 117 rxrpc_queue_call(call); 118 } else { 119 unsigned long now = jiffies, ack_at; 120 121 if (call->peer->rtt_usage > 0) 122 ack_at = nsecs_to_jiffies(call->peer->rtt); 123 else 124 ack_at = expiry; 125 126 ack_at += now; 127 if (time_before(ack_at, call->ack_at)) { 128 WRITE_ONCE(call->ack_at, ack_at); 129 rxrpc_reduce_call_timer(call, ack_at, now, 130 rxrpc_timer_set_for_ack); 131 } 132 } 133 134 trace: 135 trace_rxrpc_propose_ack(call, why, ack_reason, serial, immediate, 136 background, outcome); 137 } 138 139 /* 140 * propose an ACK be sent, locking the call structure 141 */ 142 void rxrpc_propose_ACK(struct rxrpc_call *call, u8 ack_reason, 143 u16 skew, u32 serial, bool immediate, bool background, 144 enum rxrpc_propose_ack_trace why) 145 { 146 spin_lock_bh(&call->lock); 147 __rxrpc_propose_ACK(call, ack_reason, skew, serial, 148 immediate, background, why); 149 spin_unlock_bh(&call->lock); 150 } 151 152 /* 153 * Handle congestion being detected by the retransmit timeout. 154 */ 155 static void rxrpc_congestion_timeout(struct rxrpc_call *call) 156 { 157 set_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags); 158 } 159 160 /* 161 * Perform retransmission of NAK'd and unack'd packets. 162 */ 163 static void rxrpc_resend(struct rxrpc_call *call, unsigned long now_j) 164 { 165 struct rxrpc_skb_priv *sp; 166 struct sk_buff *skb; 167 unsigned long resend_at; 168 rxrpc_seq_t cursor, seq, top; 169 ktime_t now, max_age, oldest, ack_ts, timeout, min_timeo; 170 int ix; 171 u8 annotation, anno_type, retrans = 0, unacked = 0; 172 173 _enter("{%d,%d}", call->tx_hard_ack, call->tx_top); 174 175 if (call->peer->rtt_usage > 1) 176 timeout = ns_to_ktime(call->peer->rtt * 3 / 2); 177 else 178 timeout = ms_to_ktime(rxrpc_resend_timeout); 179 min_timeo = ns_to_ktime((1000000000 / HZ) * 4); 180 if (ktime_before(timeout, min_timeo)) 181 timeout = min_timeo; 182 183 now = ktime_get_real(); 184 max_age = ktime_sub(now, timeout); 185 186 spin_lock_bh(&call->lock); 187 188 cursor = call->tx_hard_ack; 189 top = call->tx_top; 190 ASSERT(before_eq(cursor, top)); 191 if (cursor == top) 192 goto out_unlock; 193 194 /* Scan the packet list without dropping the lock and decide which of 195 * the packets in the Tx buffer we're going to resend and what the new 196 * resend timeout will be. 197 */ 198 trace_rxrpc_resend(call, (cursor + 1) & RXRPC_RXTX_BUFF_MASK); 199 oldest = now; 200 for (seq = cursor + 1; before_eq(seq, top); seq++) { 201 ix = seq & RXRPC_RXTX_BUFF_MASK; 202 annotation = call->rxtx_annotations[ix]; 203 anno_type = annotation & RXRPC_TX_ANNO_MASK; 204 annotation &= ~RXRPC_TX_ANNO_MASK; 205 if (anno_type == RXRPC_TX_ANNO_ACK) 206 continue; 207 208 skb = call->rxtx_buffer[ix]; 209 rxrpc_see_skb(skb, rxrpc_skb_tx_seen); 210 sp = rxrpc_skb(skb); 211 212 if (anno_type == RXRPC_TX_ANNO_UNACK) { 213 if (ktime_after(skb->tstamp, max_age)) { 214 if (ktime_before(skb->tstamp, oldest)) 215 oldest = skb->tstamp; 216 continue; 217 } 218 if (!(annotation & RXRPC_TX_ANNO_RESENT)) 219 unacked++; 220 } 221 222 /* Okay, we need to retransmit a packet. */ 223 call->rxtx_annotations[ix] = RXRPC_TX_ANNO_RETRANS | annotation; 224 retrans++; 225 trace_rxrpc_retransmit(call, seq, annotation | anno_type, 226 ktime_to_ns(ktime_sub(skb->tstamp, max_age))); 227 } 228 229 resend_at = nsecs_to_jiffies(ktime_to_ns(ktime_sub(now, oldest))); 230 resend_at += jiffies + rxrpc_resend_timeout; 231 WRITE_ONCE(call->resend_at, resend_at); 232 233 if (unacked) 234 rxrpc_congestion_timeout(call); 235 236 /* If there was nothing that needed retransmission then it's likely 237 * that an ACK got lost somewhere. Send a ping to find out instead of 238 * retransmitting data. 239 */ 240 if (!retrans) { 241 rxrpc_reduce_call_timer(call, resend_at, now_j, 242 rxrpc_timer_set_for_resend); 243 spin_unlock_bh(&call->lock); 244 ack_ts = ktime_sub(now, call->acks_latest_ts); 245 if (ktime_to_ns(ack_ts) < call->peer->rtt) 246 goto out; 247 rxrpc_propose_ACK(call, RXRPC_ACK_PING, 0, 0, true, false, 248 rxrpc_propose_ack_ping_for_lost_ack); 249 rxrpc_send_ack_packet(call, true, NULL); 250 goto out; 251 } 252 253 /* Now go through the Tx window and perform the retransmissions. We 254 * have to drop the lock for each send. If an ACK comes in whilst the 255 * lock is dropped, it may clear some of the retransmission markers for 256 * packets that it soft-ACKs. 257 */ 258 for (seq = cursor + 1; before_eq(seq, top); seq++) { 259 ix = seq & RXRPC_RXTX_BUFF_MASK; 260 annotation = call->rxtx_annotations[ix]; 261 anno_type = annotation & RXRPC_TX_ANNO_MASK; 262 if (anno_type != RXRPC_TX_ANNO_RETRANS) 263 continue; 264 265 skb = call->rxtx_buffer[ix]; 266 rxrpc_get_skb(skb, rxrpc_skb_tx_got); 267 spin_unlock_bh(&call->lock); 268 269 if (rxrpc_send_data_packet(call, skb, true) < 0) { 270 rxrpc_free_skb(skb, rxrpc_skb_tx_freed); 271 return; 272 } 273 274 if (rxrpc_is_client_call(call)) 275 rxrpc_expose_client_call(call); 276 277 rxrpc_free_skb(skb, rxrpc_skb_tx_freed); 278 spin_lock_bh(&call->lock); 279 280 /* We need to clear the retransmit state, but there are two 281 * things we need to be aware of: A new ACK/NAK might have been 282 * received and the packet might have been hard-ACK'd (in which 283 * case it will no longer be in the buffer). 284 */ 285 if (after(seq, call->tx_hard_ack)) { 286 annotation = call->rxtx_annotations[ix]; 287 anno_type = annotation & RXRPC_TX_ANNO_MASK; 288 if (anno_type == RXRPC_TX_ANNO_RETRANS || 289 anno_type == RXRPC_TX_ANNO_NAK) { 290 annotation &= ~RXRPC_TX_ANNO_MASK; 291 annotation |= RXRPC_TX_ANNO_UNACK; 292 } 293 annotation |= RXRPC_TX_ANNO_RESENT; 294 call->rxtx_annotations[ix] = annotation; 295 } 296 297 if (after(call->tx_hard_ack, seq)) 298 seq = call->tx_hard_ack; 299 } 300 301 out_unlock: 302 spin_unlock_bh(&call->lock); 303 out: 304 _leave(""); 305 } 306 307 /* 308 * Handle retransmission and deferred ACK/abort generation. 309 */ 310 void rxrpc_process_call(struct work_struct *work) 311 { 312 struct rxrpc_call *call = 313 container_of(work, struct rxrpc_call, processor); 314 rxrpc_serial_t *send_ack; 315 unsigned long now, next, t; 316 317 rxrpc_see_call(call); 318 319 //printk("\n--------------------\n"); 320 _enter("{%d,%s,%lx}", 321 call->debug_id, rxrpc_call_states[call->state], call->events); 322 323 recheck_state: 324 if (test_and_clear_bit(RXRPC_CALL_EV_ABORT, &call->events)) { 325 rxrpc_send_abort_packet(call); 326 goto recheck_state; 327 } 328 329 if (call->state == RXRPC_CALL_COMPLETE) { 330 del_timer_sync(&call->timer); 331 rxrpc_notify_socket(call); 332 goto out_put; 333 } 334 335 /* Work out if any timeouts tripped */ 336 now = jiffies; 337 t = READ_ONCE(call->expect_rx_by); 338 if (time_after_eq(now, t)) { 339 trace_rxrpc_timer(call, rxrpc_timer_exp_normal, now); 340 set_bit(RXRPC_CALL_EV_EXPIRED, &call->events); 341 } 342 343 t = READ_ONCE(call->expect_req_by); 344 if (call->state == RXRPC_CALL_SERVER_RECV_REQUEST && 345 time_after_eq(now, t)) { 346 trace_rxrpc_timer(call, rxrpc_timer_exp_idle, now); 347 set_bit(RXRPC_CALL_EV_EXPIRED, &call->events); 348 } 349 350 t = READ_ONCE(call->expect_term_by); 351 if (time_after_eq(now, t)) { 352 trace_rxrpc_timer(call, rxrpc_timer_exp_hard, now); 353 set_bit(RXRPC_CALL_EV_EXPIRED, &call->events); 354 } 355 356 t = READ_ONCE(call->ack_at); 357 if (time_after_eq(now, t)) { 358 trace_rxrpc_timer(call, rxrpc_timer_exp_ack, now); 359 cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET); 360 set_bit(RXRPC_CALL_EV_ACK, &call->events); 361 } 362 363 t = READ_ONCE(call->ack_lost_at); 364 if (time_after_eq(now, t)) { 365 trace_rxrpc_timer(call, rxrpc_timer_exp_lost_ack, now); 366 cmpxchg(&call->ack_lost_at, t, now + MAX_JIFFY_OFFSET); 367 set_bit(RXRPC_CALL_EV_ACK_LOST, &call->events); 368 } 369 370 t = READ_ONCE(call->keepalive_at); 371 if (time_after_eq(now, t)) { 372 trace_rxrpc_timer(call, rxrpc_timer_exp_keepalive, now); 373 cmpxchg(&call->keepalive_at, t, now + MAX_JIFFY_OFFSET); 374 rxrpc_propose_ACK(call, RXRPC_ACK_PING, 0, 0, true, true, 375 rxrpc_propose_ack_ping_for_keepalive); 376 set_bit(RXRPC_CALL_EV_PING, &call->events); 377 } 378 379 t = READ_ONCE(call->ping_at); 380 if (time_after_eq(now, t)) { 381 trace_rxrpc_timer(call, rxrpc_timer_exp_ping, now); 382 cmpxchg(&call->ping_at, t, now + MAX_JIFFY_OFFSET); 383 set_bit(RXRPC_CALL_EV_PING, &call->events); 384 } 385 386 t = READ_ONCE(call->resend_at); 387 if (time_after_eq(now, t)) { 388 trace_rxrpc_timer(call, rxrpc_timer_exp_resend, now); 389 cmpxchg(&call->resend_at, t, now + MAX_JIFFY_OFFSET); 390 set_bit(RXRPC_CALL_EV_RESEND, &call->events); 391 } 392 393 /* Process events */ 394 if (test_and_clear_bit(RXRPC_CALL_EV_EXPIRED, &call->events)) { 395 if (test_bit(RXRPC_CALL_RX_HEARD, &call->flags) && 396 (int)call->conn->hi_serial - (int)call->rx_serial > 0) { 397 trace_rxrpc_call_reset(call); 398 rxrpc_abort_call("EXP", call, 0, RX_USER_ABORT, -ECONNRESET); 399 } else { 400 rxrpc_abort_call("EXP", call, 0, RX_USER_ABORT, -ETIME); 401 } 402 set_bit(RXRPC_CALL_EV_ABORT, &call->events); 403 goto recheck_state; 404 } 405 406 send_ack = NULL; 407 if (test_and_clear_bit(RXRPC_CALL_EV_ACK_LOST, &call->events)) { 408 call->acks_lost_top = call->tx_top; 409 rxrpc_propose_ACK(call, RXRPC_ACK_PING, 0, 0, true, false, 410 rxrpc_propose_ack_ping_for_lost_ack); 411 send_ack = &call->acks_lost_ping; 412 } 413 414 if (test_and_clear_bit(RXRPC_CALL_EV_ACK, &call->events) || 415 send_ack) { 416 if (call->ackr_reason) { 417 rxrpc_send_ack_packet(call, false, send_ack); 418 goto recheck_state; 419 } 420 } 421 422 if (test_and_clear_bit(RXRPC_CALL_EV_PING, &call->events)) { 423 rxrpc_send_ack_packet(call, true, NULL); 424 goto recheck_state; 425 } 426 427 if (test_and_clear_bit(RXRPC_CALL_EV_RESEND, &call->events)) { 428 rxrpc_resend(call, now); 429 goto recheck_state; 430 } 431 432 /* Make sure the timer is restarted */ 433 next = call->expect_rx_by; 434 435 #define set(T) { t = READ_ONCE(T); if (time_before(t, next)) next = t; } 436 437 set(call->expect_req_by); 438 set(call->expect_term_by); 439 set(call->ack_at); 440 set(call->ack_lost_at); 441 set(call->resend_at); 442 set(call->keepalive_at); 443 set(call->ping_at); 444 445 now = jiffies; 446 if (time_after_eq(now, next)) 447 goto recheck_state; 448 449 rxrpc_reduce_call_timer(call, next, now, rxrpc_timer_restart); 450 451 /* other events may have been raised since we started checking */ 452 if (call->events && call->state < RXRPC_CALL_COMPLETE) { 453 __rxrpc_queue_call(call); 454 goto out; 455 } 456 457 out_put: 458 rxrpc_put_call(call, rxrpc_call_put); 459 out: 460 _leave(""); 461 } 462