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