xref: /openbmc/linux/net/rxrpc/input.c (revision f0009360)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* Processing of received RxRPC packets
3  *
4  * Copyright (C) 2020 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 "ar-internal.h"
11 
12 static void rxrpc_proto_abort(const char *why,
13 			      struct rxrpc_call *call, rxrpc_seq_t seq)
14 {
15 	if (rxrpc_abort_call(why, call, seq, RX_PROTOCOL_ERROR, -EBADMSG))
16 		rxrpc_send_abort_packet(call);
17 }
18 
19 /*
20  * Do TCP-style congestion management [RFC 5681].
21  */
22 static void rxrpc_congestion_management(struct rxrpc_call *call,
23 					struct sk_buff *skb,
24 					struct rxrpc_ack_summary *summary,
25 					rxrpc_serial_t acked_serial)
26 {
27 	enum rxrpc_congest_change change = rxrpc_cong_no_change;
28 	unsigned int cumulative_acks = call->cong_cumul_acks;
29 	unsigned int cwnd = call->cong_cwnd;
30 	bool resend = false;
31 
32 	summary->flight_size =
33 		(call->tx_top - call->acks_hard_ack) - summary->nr_acks;
34 
35 	if (test_and_clear_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags)) {
36 		summary->retrans_timeo = true;
37 		call->cong_ssthresh = max_t(unsigned int,
38 					    summary->flight_size / 2, 2);
39 		cwnd = 1;
40 		if (cwnd >= call->cong_ssthresh &&
41 		    call->cong_mode == RXRPC_CALL_SLOW_START) {
42 			call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
43 			call->cong_tstamp = skb->tstamp;
44 			cumulative_acks = 0;
45 		}
46 	}
47 
48 	cumulative_acks += summary->nr_new_acks;
49 	cumulative_acks += summary->nr_rot_new_acks;
50 	if (cumulative_acks > 255)
51 		cumulative_acks = 255;
52 
53 	summary->mode = call->cong_mode;
54 	summary->cwnd = call->cong_cwnd;
55 	summary->ssthresh = call->cong_ssthresh;
56 	summary->cumulative_acks = cumulative_acks;
57 	summary->dup_acks = call->cong_dup_acks;
58 
59 	switch (call->cong_mode) {
60 	case RXRPC_CALL_SLOW_START:
61 		if (summary->saw_nacks)
62 			goto packet_loss_detected;
63 		if (summary->cumulative_acks > 0)
64 			cwnd += 1;
65 		if (cwnd >= call->cong_ssthresh) {
66 			call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
67 			call->cong_tstamp = skb->tstamp;
68 		}
69 		goto out;
70 
71 	case RXRPC_CALL_CONGEST_AVOIDANCE:
72 		if (summary->saw_nacks)
73 			goto packet_loss_detected;
74 
75 		/* We analyse the number of packets that get ACK'd per RTT
76 		 * period and increase the window if we managed to fill it.
77 		 */
78 		if (call->peer->rtt_count == 0)
79 			goto out;
80 		if (ktime_before(skb->tstamp,
81 				 ktime_add_us(call->cong_tstamp,
82 					      call->peer->srtt_us >> 3)))
83 			goto out_no_clear_ca;
84 		change = rxrpc_cong_rtt_window_end;
85 		call->cong_tstamp = skb->tstamp;
86 		if (cumulative_acks >= cwnd)
87 			cwnd++;
88 		goto out;
89 
90 	case RXRPC_CALL_PACKET_LOSS:
91 		if (!summary->saw_nacks)
92 			goto resume_normality;
93 
94 		if (summary->new_low_nack) {
95 			change = rxrpc_cong_new_low_nack;
96 			call->cong_dup_acks = 1;
97 			if (call->cong_extra > 1)
98 				call->cong_extra = 1;
99 			goto send_extra_data;
100 		}
101 
102 		call->cong_dup_acks++;
103 		if (call->cong_dup_acks < 3)
104 			goto send_extra_data;
105 
106 		change = rxrpc_cong_begin_retransmission;
107 		call->cong_mode = RXRPC_CALL_FAST_RETRANSMIT;
108 		call->cong_ssthresh = max_t(unsigned int,
109 					    summary->flight_size / 2, 2);
110 		cwnd = call->cong_ssthresh + 3;
111 		call->cong_extra = 0;
112 		call->cong_dup_acks = 0;
113 		resend = true;
114 		goto out;
115 
116 	case RXRPC_CALL_FAST_RETRANSMIT:
117 		if (!summary->new_low_nack) {
118 			if (summary->nr_new_acks == 0)
119 				cwnd += 1;
120 			call->cong_dup_acks++;
121 			if (call->cong_dup_acks == 2) {
122 				change = rxrpc_cong_retransmit_again;
123 				call->cong_dup_acks = 0;
124 				resend = true;
125 			}
126 		} else {
127 			change = rxrpc_cong_progress;
128 			cwnd = call->cong_ssthresh;
129 			if (!summary->saw_nacks)
130 				goto resume_normality;
131 		}
132 		goto out;
133 
134 	default:
135 		BUG();
136 		goto out;
137 	}
138 
139 resume_normality:
140 	change = rxrpc_cong_cleared_nacks;
141 	call->cong_dup_acks = 0;
142 	call->cong_extra = 0;
143 	call->cong_tstamp = skb->tstamp;
144 	if (cwnd < call->cong_ssthresh)
145 		call->cong_mode = RXRPC_CALL_SLOW_START;
146 	else
147 		call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
148 out:
149 	cumulative_acks = 0;
150 out_no_clear_ca:
151 	if (cwnd >= RXRPC_TX_MAX_WINDOW)
152 		cwnd = RXRPC_TX_MAX_WINDOW;
153 	call->cong_cwnd = cwnd;
154 	call->cong_cumul_acks = cumulative_acks;
155 	trace_rxrpc_congest(call, summary, acked_serial, change);
156 	if (resend)
157 		rxrpc_resend(call, skb);
158 	return;
159 
160 packet_loss_detected:
161 	change = rxrpc_cong_saw_nack;
162 	call->cong_mode = RXRPC_CALL_PACKET_LOSS;
163 	call->cong_dup_acks = 0;
164 	goto send_extra_data;
165 
166 send_extra_data:
167 	/* Send some previously unsent DATA if we have some to advance the ACK
168 	 * state.
169 	 */
170 	if (test_bit(RXRPC_CALL_TX_LAST, &call->flags) ||
171 	    summary->nr_acks != call->tx_top - call->acks_hard_ack) {
172 		call->cong_extra++;
173 		wake_up(&call->waitq);
174 	}
175 	goto out_no_clear_ca;
176 }
177 
178 /*
179  * Degrade the congestion window if we haven't transmitted a packet for >1RTT.
180  */
181 void rxrpc_congestion_degrade(struct rxrpc_call *call)
182 {
183 	ktime_t rtt, now;
184 
185 	if (call->cong_mode != RXRPC_CALL_SLOW_START &&
186 	    call->cong_mode != RXRPC_CALL_CONGEST_AVOIDANCE)
187 		return;
188 	if (call->state == RXRPC_CALL_CLIENT_AWAIT_REPLY)
189 		return;
190 
191 	rtt = ns_to_ktime(call->peer->srtt_us * (1000 / 8));
192 	now = ktime_get_real();
193 	if (!ktime_before(ktime_add(call->tx_last_sent, rtt), now))
194 		return;
195 
196 	trace_rxrpc_reset_cwnd(call, now);
197 	rxrpc_inc_stat(call->rxnet, stat_tx_data_cwnd_reset);
198 	call->tx_last_sent = now;
199 	call->cong_mode = RXRPC_CALL_SLOW_START;
200 	call->cong_ssthresh = max_t(unsigned int, call->cong_ssthresh,
201 				    call->cong_cwnd * 3 / 4);
202 	call->cong_cwnd = max_t(unsigned int, call->cong_cwnd / 2, RXRPC_MIN_CWND);
203 }
204 
205 /*
206  * Apply a hard ACK by advancing the Tx window.
207  */
208 static bool rxrpc_rotate_tx_window(struct rxrpc_call *call, rxrpc_seq_t to,
209 				   struct rxrpc_ack_summary *summary)
210 {
211 	struct rxrpc_txbuf *txb;
212 	bool rot_last = false;
213 
214 	list_for_each_entry_rcu(txb, &call->tx_buffer, call_link, false) {
215 		if (before_eq(txb->seq, call->acks_hard_ack))
216 			continue;
217 		summary->nr_rot_new_acks++;
218 		if (test_bit(RXRPC_TXBUF_LAST, &txb->flags)) {
219 			set_bit(RXRPC_CALL_TX_LAST, &call->flags);
220 			rot_last = true;
221 		}
222 		if (txb->seq == to)
223 			break;
224 	}
225 
226 	if (rot_last)
227 		set_bit(RXRPC_CALL_TX_ALL_ACKED, &call->flags);
228 
229 	_enter("%x,%x,%x,%d", to, call->acks_hard_ack, call->tx_top, rot_last);
230 
231 	if (call->acks_lowest_nak == call->acks_hard_ack) {
232 		call->acks_lowest_nak = to;
233 	} else if (after(to, call->acks_lowest_nak)) {
234 		summary->new_low_nack = true;
235 		call->acks_lowest_nak = to;
236 	}
237 
238 	smp_store_release(&call->acks_hard_ack, to);
239 
240 	trace_rxrpc_txqueue(call, (rot_last ?
241 				   rxrpc_txqueue_rotate_last :
242 				   rxrpc_txqueue_rotate));
243 	wake_up(&call->waitq);
244 	return rot_last;
245 }
246 
247 /*
248  * End the transmission phase of a call.
249  *
250  * This occurs when we get an ACKALL packet, the first DATA packet of a reply,
251  * or a final ACK packet.
252  */
253 static bool rxrpc_end_tx_phase(struct rxrpc_call *call, bool reply_begun,
254 			       const char *abort_why)
255 {
256 	unsigned int state;
257 
258 	ASSERT(test_bit(RXRPC_CALL_TX_LAST, &call->flags));
259 
260 	write_lock(&call->state_lock);
261 
262 	state = call->state;
263 	switch (state) {
264 	case RXRPC_CALL_CLIENT_SEND_REQUEST:
265 	case RXRPC_CALL_CLIENT_AWAIT_REPLY:
266 		if (reply_begun)
267 			call->state = state = RXRPC_CALL_CLIENT_RECV_REPLY;
268 		else
269 			call->state = state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
270 		break;
271 
272 	case RXRPC_CALL_SERVER_AWAIT_ACK:
273 		__rxrpc_call_completed(call);
274 		state = call->state;
275 		break;
276 
277 	default:
278 		goto bad_state;
279 	}
280 
281 	write_unlock(&call->state_lock);
282 	if (state == RXRPC_CALL_CLIENT_AWAIT_REPLY)
283 		trace_rxrpc_txqueue(call, rxrpc_txqueue_await_reply);
284 	else
285 		trace_rxrpc_txqueue(call, rxrpc_txqueue_end);
286 	_leave(" = ok");
287 	return true;
288 
289 bad_state:
290 	write_unlock(&call->state_lock);
291 	kdebug("end_tx %s", rxrpc_call_states[call->state]);
292 	rxrpc_proto_abort(abort_why, call, call->tx_top);
293 	return false;
294 }
295 
296 /*
297  * Begin the reply reception phase of a call.
298  */
299 static bool rxrpc_receiving_reply(struct rxrpc_call *call)
300 {
301 	struct rxrpc_ack_summary summary = { 0 };
302 	unsigned long now, timo;
303 	rxrpc_seq_t top = READ_ONCE(call->tx_top);
304 
305 	if (call->ackr_reason) {
306 		now = jiffies;
307 		timo = now + MAX_JIFFY_OFFSET;
308 		WRITE_ONCE(call->resend_at, timo);
309 		WRITE_ONCE(call->delay_ack_at, timo);
310 		trace_rxrpc_timer(call, rxrpc_timer_init_for_reply, now);
311 	}
312 
313 	if (!test_bit(RXRPC_CALL_TX_LAST, &call->flags)) {
314 		if (!rxrpc_rotate_tx_window(call, top, &summary)) {
315 			rxrpc_proto_abort("TXL", call, top);
316 			return false;
317 		}
318 	}
319 	return rxrpc_end_tx_phase(call, true, "ETD");
320 }
321 
322 static void rxrpc_input_update_ack_window(struct rxrpc_call *call,
323 					  rxrpc_seq_t window, rxrpc_seq_t wtop)
324 {
325 	atomic64_set_release(&call->ackr_window, ((u64)wtop) << 32 | window);
326 }
327 
328 /*
329  * Push a DATA packet onto the Rx queue.
330  */
331 static void rxrpc_input_queue_data(struct rxrpc_call *call, struct sk_buff *skb,
332 				   rxrpc_seq_t window, rxrpc_seq_t wtop,
333 				   enum rxrpc_receive_trace why)
334 {
335 	struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
336 	bool last = sp->hdr.flags & RXRPC_LAST_PACKET;
337 
338 	__skb_queue_tail(&call->recvmsg_queue, skb);
339 	rxrpc_input_update_ack_window(call, window, wtop);
340 
341 	trace_rxrpc_receive(call, last ? why + 1 : why, sp->hdr.serial, sp->hdr.seq);
342 }
343 
344 /*
345  * Process a DATA packet.
346  */
347 static void rxrpc_input_data_one(struct rxrpc_call *call, struct sk_buff *skb,
348 				 bool *_notify)
349 {
350 	struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
351 	struct sk_buff *oos;
352 	rxrpc_serial_t serial = sp->hdr.serial;
353 	u64 win = atomic64_read(&call->ackr_window);
354 	rxrpc_seq_t window = lower_32_bits(win);
355 	rxrpc_seq_t wtop = upper_32_bits(win);
356 	rxrpc_seq_t wlimit = window + call->rx_winsize - 1;
357 	rxrpc_seq_t seq = sp->hdr.seq;
358 	bool last = sp->hdr.flags & RXRPC_LAST_PACKET;
359 	int ack_reason = -1;
360 
361 	rxrpc_inc_stat(call->rxnet, stat_rx_data);
362 	if (sp->hdr.flags & RXRPC_REQUEST_ACK)
363 		rxrpc_inc_stat(call->rxnet, stat_rx_data_reqack);
364 	if (sp->hdr.flags & RXRPC_JUMBO_PACKET)
365 		rxrpc_inc_stat(call->rxnet, stat_rx_data_jumbo);
366 
367 	if (last) {
368 		if (test_and_set_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
369 		    seq + 1 != wtop) {
370 			rxrpc_proto_abort("LSN", call, seq);
371 			return;
372 		}
373 	} else {
374 		if (test_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
375 		    after_eq(seq, wtop)) {
376 			pr_warn("Packet beyond last: c=%x q=%x window=%x-%x wlimit=%x\n",
377 				call->debug_id, seq, window, wtop, wlimit);
378 			rxrpc_proto_abort("LSA", call, seq);
379 			return;
380 		}
381 	}
382 
383 	if (after(seq, call->rx_highest_seq))
384 		call->rx_highest_seq = seq;
385 
386 	trace_rxrpc_rx_data(call->debug_id, seq, serial, sp->hdr.flags);
387 
388 	if (before(seq, window)) {
389 		ack_reason = RXRPC_ACK_DUPLICATE;
390 		goto send_ack;
391 	}
392 	if (after(seq, wlimit)) {
393 		ack_reason = RXRPC_ACK_EXCEEDS_WINDOW;
394 		goto send_ack;
395 	}
396 
397 	/* Queue the packet. */
398 	if (seq == window) {
399 		rxrpc_seq_t reset_from;
400 		bool reset_sack = false;
401 
402 		if (sp->hdr.flags & RXRPC_REQUEST_ACK)
403 			ack_reason = RXRPC_ACK_REQUESTED;
404 		/* Send an immediate ACK if we fill in a hole */
405 		else if (!skb_queue_empty(&call->rx_oos_queue))
406 			ack_reason = RXRPC_ACK_DELAY;
407 		else
408 			atomic_inc_return(&call->ackr_nr_unacked);
409 
410 		window++;
411 		if (after(window, wtop))
412 			wtop = window;
413 
414 		rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg);
415 
416 		spin_lock(&call->recvmsg_queue.lock);
417 		rxrpc_input_queue_data(call, skb, window, wtop, rxrpc_receive_queue);
418 		*_notify = true;
419 
420 		while ((oos = skb_peek(&call->rx_oos_queue))) {
421 			struct rxrpc_skb_priv *osp = rxrpc_skb(oos);
422 
423 			if (after(osp->hdr.seq, window))
424 				break;
425 
426 			__skb_unlink(oos, &call->rx_oos_queue);
427 			last = osp->hdr.flags & RXRPC_LAST_PACKET;
428 			seq = osp->hdr.seq;
429 			if (!reset_sack) {
430 				reset_from = seq;
431 				reset_sack = true;
432 			}
433 
434 			window++;
435 			rxrpc_input_queue_data(call, oos, window, wtop,
436 						 rxrpc_receive_queue_oos);
437 		}
438 
439 		spin_unlock(&call->recvmsg_queue.lock);
440 
441 		if (reset_sack) {
442 			do {
443 				call->ackr_sack_table[reset_from % RXRPC_SACK_SIZE] = 0;
444 			} while (reset_from++, before(reset_from, window));
445 		}
446 	} else {
447 		bool keep = false;
448 
449 		ack_reason = RXRPC_ACK_OUT_OF_SEQUENCE;
450 
451 		if (!call->ackr_sack_table[seq % RXRPC_SACK_SIZE]) {
452 			call->ackr_sack_table[seq % RXRPC_SACK_SIZE] = 1;
453 			keep = 1;
454 		}
455 
456 		if (after(seq + 1, wtop)) {
457 			wtop = seq + 1;
458 			rxrpc_input_update_ack_window(call, window, wtop);
459 		}
460 
461 		if (!keep) {
462 			ack_reason = RXRPC_ACK_DUPLICATE;
463 			goto send_ack;
464 		}
465 
466 		skb_queue_walk(&call->rx_oos_queue, oos) {
467 			struct rxrpc_skb_priv *osp = rxrpc_skb(oos);
468 
469 			if (after(osp->hdr.seq, seq)) {
470 				rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg_oos);
471 				__skb_queue_before(&call->rx_oos_queue, oos, skb);
472 				goto oos_queued;
473 			}
474 		}
475 
476 		rxrpc_get_skb(skb, rxrpc_skb_get_to_recvmsg_oos);
477 		__skb_queue_tail(&call->rx_oos_queue, skb);
478 	oos_queued:
479 		trace_rxrpc_receive(call, last ? rxrpc_receive_oos_last : rxrpc_receive_oos,
480 				    sp->hdr.serial, sp->hdr.seq);
481 	}
482 
483 send_ack:
484 	if (ack_reason >= 0)
485 		rxrpc_send_ACK(call, ack_reason, serial,
486 			       rxrpc_propose_ack_input_data);
487 	else
488 		rxrpc_propose_delay_ACK(call, serial,
489 					rxrpc_propose_ack_input_data);
490 }
491 
492 /*
493  * Split a jumbo packet and file the bits separately.
494  */
495 static bool rxrpc_input_split_jumbo(struct rxrpc_call *call, struct sk_buff *skb)
496 {
497 	struct rxrpc_jumbo_header jhdr;
498 	struct rxrpc_skb_priv *sp = rxrpc_skb(skb), *jsp;
499 	struct sk_buff *jskb;
500 	unsigned int offset = sizeof(struct rxrpc_wire_header);
501 	unsigned int len = skb->len - offset;
502 	bool notify = false;
503 
504 	while (sp->hdr.flags & RXRPC_JUMBO_PACKET) {
505 		if (len < RXRPC_JUMBO_SUBPKTLEN)
506 			goto protocol_error;
507 		if (sp->hdr.flags & RXRPC_LAST_PACKET)
508 			goto protocol_error;
509 		if (skb_copy_bits(skb, offset + RXRPC_JUMBO_DATALEN,
510 				  &jhdr, sizeof(jhdr)) < 0)
511 			goto protocol_error;
512 
513 		jskb = skb_clone(skb, GFP_NOFS);
514 		if (!jskb) {
515 			kdebug("couldn't clone");
516 			return false;
517 		}
518 		rxrpc_new_skb(jskb, rxrpc_skb_new_jumbo_subpacket);
519 		jsp = rxrpc_skb(jskb);
520 		jsp->offset = offset;
521 		jsp->len = RXRPC_JUMBO_DATALEN;
522 		rxrpc_input_data_one(call, jskb, &notify);
523 		rxrpc_free_skb(jskb, rxrpc_skb_put_jumbo_subpacket);
524 
525 		sp->hdr.flags = jhdr.flags;
526 		sp->hdr._rsvd = ntohs(jhdr._rsvd);
527 		sp->hdr.seq++;
528 		sp->hdr.serial++;
529 		offset += RXRPC_JUMBO_SUBPKTLEN;
530 		len -= RXRPC_JUMBO_SUBPKTLEN;
531 	}
532 
533 	sp->offset = offset;
534 	sp->len    = len;
535 	rxrpc_input_data_one(call, skb, &notify);
536 	if (notify) {
537 		trace_rxrpc_notify_socket(call->debug_id, sp->hdr.serial);
538 		rxrpc_notify_socket(call);
539 	}
540 	return true;
541 
542 protocol_error:
543 	return false;
544 }
545 
546 /*
547  * Process a DATA packet, adding the packet to the Rx ring.  The caller's
548  * packet ref must be passed on or discarded.
549  */
550 static void rxrpc_input_data(struct rxrpc_call *call, struct sk_buff *skb)
551 {
552 	struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
553 	enum rxrpc_call_state state;
554 	rxrpc_serial_t serial = sp->hdr.serial;
555 	rxrpc_seq_t seq0 = sp->hdr.seq;
556 
557 	_enter("{%llx,%x},{%u,%x}",
558 	       atomic64_read(&call->ackr_window), call->rx_highest_seq,
559 	       skb->len, seq0);
560 
561 	state = READ_ONCE(call->state);
562 	if (state >= RXRPC_CALL_COMPLETE)
563 		return;
564 
565 	if (state == RXRPC_CALL_SERVER_RECV_REQUEST) {
566 		unsigned long timo = READ_ONCE(call->next_req_timo);
567 		unsigned long now, expect_req_by;
568 
569 		if (timo) {
570 			now = jiffies;
571 			expect_req_by = now + timo;
572 			WRITE_ONCE(call->expect_req_by, expect_req_by);
573 			rxrpc_reduce_call_timer(call, expect_req_by, now,
574 						rxrpc_timer_set_for_idle);
575 		}
576 	}
577 
578 	/* Received data implicitly ACKs all of the request packets we sent
579 	 * when we're acting as a client.
580 	 */
581 	if ((state == RXRPC_CALL_CLIENT_SEND_REQUEST ||
582 	     state == RXRPC_CALL_CLIENT_AWAIT_REPLY) &&
583 	    !rxrpc_receiving_reply(call))
584 		goto out_notify;
585 
586 	if (!rxrpc_input_split_jumbo(call, skb)) {
587 		rxrpc_proto_abort("VLD", call, sp->hdr.seq);
588 		goto out_notify;
589 	}
590 	skb = NULL;
591 
592 out_notify:
593 	trace_rxrpc_notify_socket(call->debug_id, serial);
594 	rxrpc_notify_socket(call);
595 	_leave(" [queued]");
596 }
597 
598 /*
599  * See if there's a cached RTT probe to complete.
600  */
601 static void rxrpc_complete_rtt_probe(struct rxrpc_call *call,
602 				     ktime_t resp_time,
603 				     rxrpc_serial_t acked_serial,
604 				     rxrpc_serial_t ack_serial,
605 				     enum rxrpc_rtt_rx_trace type)
606 {
607 	rxrpc_serial_t orig_serial;
608 	unsigned long avail;
609 	ktime_t sent_at;
610 	bool matched = false;
611 	int i;
612 
613 	avail = READ_ONCE(call->rtt_avail);
614 	smp_rmb(); /* Read avail bits before accessing data. */
615 
616 	for (i = 0; i < ARRAY_SIZE(call->rtt_serial); i++) {
617 		if (!test_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &avail))
618 			continue;
619 
620 		sent_at = call->rtt_sent_at[i];
621 		orig_serial = call->rtt_serial[i];
622 
623 		if (orig_serial == acked_serial) {
624 			clear_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
625 			smp_mb(); /* Read data before setting avail bit */
626 			set_bit(i, &call->rtt_avail);
627 			if (type != rxrpc_rtt_rx_cancel)
628 				rxrpc_peer_add_rtt(call, type, i, acked_serial, ack_serial,
629 						   sent_at, resp_time);
630 			else
631 				trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_cancel, i,
632 						   orig_serial, acked_serial, 0, 0);
633 			matched = true;
634 		}
635 
636 		/* If a later serial is being acked, then mark this slot as
637 		 * being available.
638 		 */
639 		if (after(acked_serial, orig_serial)) {
640 			trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_obsolete, i,
641 					   orig_serial, acked_serial, 0, 0);
642 			clear_bit(i + RXRPC_CALL_RTT_PEND_SHIFT, &call->rtt_avail);
643 			smp_wmb();
644 			set_bit(i, &call->rtt_avail);
645 		}
646 	}
647 
648 	if (!matched)
649 		trace_rxrpc_rtt_rx(call, rxrpc_rtt_rx_lost, 9, 0, acked_serial, 0, 0);
650 }
651 
652 /*
653  * Process the extra information that may be appended to an ACK packet
654  */
655 static void rxrpc_input_ackinfo(struct rxrpc_call *call, struct sk_buff *skb,
656 				struct rxrpc_ackinfo *ackinfo)
657 {
658 	struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
659 	struct rxrpc_peer *peer;
660 	unsigned int mtu;
661 	bool wake = false;
662 	u32 rwind = ntohl(ackinfo->rwind);
663 
664 	if (rwind > RXRPC_TX_MAX_WINDOW)
665 		rwind = RXRPC_TX_MAX_WINDOW;
666 	if (call->tx_winsize != rwind) {
667 		if (rwind > call->tx_winsize)
668 			wake = true;
669 		trace_rxrpc_rx_rwind_change(call, sp->hdr.serial, rwind, wake);
670 		call->tx_winsize = rwind;
671 	}
672 
673 	if (call->cong_ssthresh > rwind)
674 		call->cong_ssthresh = rwind;
675 
676 	mtu = min(ntohl(ackinfo->rxMTU), ntohl(ackinfo->maxMTU));
677 
678 	peer = call->peer;
679 	if (mtu < peer->maxdata) {
680 		spin_lock(&peer->lock);
681 		peer->maxdata = mtu;
682 		peer->mtu = mtu + peer->hdrsize;
683 		spin_unlock(&peer->lock);
684 	}
685 
686 	if (wake)
687 		wake_up(&call->waitq);
688 }
689 
690 /*
691  * Process individual soft ACKs.
692  *
693  * Each ACK in the array corresponds to one packet and can be either an ACK or
694  * a NAK.  If we get find an explicitly NAK'd packet we resend immediately;
695  * packets that lie beyond the end of the ACK list are scheduled for resend by
696  * the timer on the basis that the peer might just not have processed them at
697  * the time the ACK was sent.
698  */
699 static void rxrpc_input_soft_acks(struct rxrpc_call *call, u8 *acks,
700 				  rxrpc_seq_t seq, int nr_acks,
701 				  struct rxrpc_ack_summary *summary)
702 {
703 	unsigned int i;
704 
705 	for (i = 0; i < nr_acks; i++) {
706 		if (acks[i] == RXRPC_ACK_TYPE_ACK) {
707 			summary->nr_acks++;
708 			summary->nr_new_acks++;
709 		} else {
710 			if (!summary->saw_nacks &&
711 			    call->acks_lowest_nak != seq + i) {
712 				call->acks_lowest_nak = seq + i;
713 				summary->new_low_nack = true;
714 			}
715 			summary->saw_nacks = true;
716 		}
717 	}
718 }
719 
720 /*
721  * Return true if the ACK is valid - ie. it doesn't appear to have regressed
722  * with respect to the ack state conveyed by preceding ACKs.
723  */
724 static bool rxrpc_is_ack_valid(struct rxrpc_call *call,
725 			       rxrpc_seq_t first_pkt, rxrpc_seq_t prev_pkt)
726 {
727 	rxrpc_seq_t base = READ_ONCE(call->acks_first_seq);
728 
729 	if (after(first_pkt, base))
730 		return true; /* The window advanced */
731 
732 	if (before(first_pkt, base))
733 		return false; /* firstPacket regressed */
734 
735 	if (after_eq(prev_pkt, call->acks_prev_seq))
736 		return true; /* previousPacket hasn't regressed. */
737 
738 	/* Some rx implementations put a serial number in previousPacket. */
739 	if (after_eq(prev_pkt, base + call->tx_winsize))
740 		return false;
741 	return true;
742 }
743 
744 /*
745  * Process an ACK packet.
746  *
747  * ack.firstPacket is the sequence number of the first soft-ACK'd/NAK'd packet
748  * in the ACK array.  Anything before that is hard-ACK'd and may be discarded.
749  *
750  * A hard-ACK means that a packet has been processed and may be discarded; a
751  * soft-ACK means that the packet may be discarded and retransmission
752  * requested.  A phase is complete when all packets are hard-ACK'd.
753  */
754 static void rxrpc_input_ack(struct rxrpc_call *call, struct sk_buff *skb)
755 {
756 	struct rxrpc_ack_summary summary = { 0 };
757 	struct rxrpc_ackpacket ack;
758 	struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
759 	struct rxrpc_ackinfo info;
760 	rxrpc_serial_t ack_serial, acked_serial;
761 	rxrpc_seq_t first_soft_ack, hard_ack, prev_pkt;
762 	int nr_acks, offset, ioffset;
763 
764 	_enter("");
765 
766 	offset = sizeof(struct rxrpc_wire_header);
767 	if (skb_copy_bits(skb, offset, &ack, sizeof(ack)) < 0)
768 		return rxrpc_proto_abort("XAK", call, 0);
769 	offset += sizeof(ack);
770 
771 	ack_serial = sp->hdr.serial;
772 	acked_serial = ntohl(ack.serial);
773 	first_soft_ack = ntohl(ack.firstPacket);
774 	prev_pkt = ntohl(ack.previousPacket);
775 	hard_ack = first_soft_ack - 1;
776 	nr_acks = ack.nAcks;
777 	summary.ack_reason = (ack.reason < RXRPC_ACK__INVALID ?
778 			      ack.reason : RXRPC_ACK__INVALID);
779 
780 	trace_rxrpc_rx_ack(call, ack_serial, acked_serial,
781 			   first_soft_ack, prev_pkt,
782 			   summary.ack_reason, nr_acks);
783 	rxrpc_inc_stat(call->rxnet, stat_rx_acks[ack.reason]);
784 
785 	switch (ack.reason) {
786 	case RXRPC_ACK_PING_RESPONSE:
787 		rxrpc_complete_rtt_probe(call, skb->tstamp, acked_serial, ack_serial,
788 					 rxrpc_rtt_rx_ping_response);
789 		break;
790 	case RXRPC_ACK_REQUESTED:
791 		rxrpc_complete_rtt_probe(call, skb->tstamp, acked_serial, ack_serial,
792 					 rxrpc_rtt_rx_requested_ack);
793 		break;
794 	default:
795 		if (acked_serial != 0)
796 			rxrpc_complete_rtt_probe(call, skb->tstamp, acked_serial, ack_serial,
797 						 rxrpc_rtt_rx_cancel);
798 		break;
799 	}
800 
801 	if (ack.reason == RXRPC_ACK_PING) {
802 		rxrpc_send_ACK(call, RXRPC_ACK_PING_RESPONSE, ack_serial,
803 			       rxrpc_propose_ack_respond_to_ping);
804 	} else if (sp->hdr.flags & RXRPC_REQUEST_ACK) {
805 		rxrpc_send_ACK(call, RXRPC_ACK_REQUESTED, ack_serial,
806 			       rxrpc_propose_ack_respond_to_ack);
807 	}
808 
809 	/* If we get an EXCEEDS_WINDOW ACK from the server, it probably
810 	 * indicates that the client address changed due to NAT.  The server
811 	 * lost the call because it switched to a different peer.
812 	 */
813 	if (unlikely(ack.reason == RXRPC_ACK_EXCEEDS_WINDOW) &&
814 	    first_soft_ack == 1 &&
815 	    prev_pkt == 0 &&
816 	    rxrpc_is_client_call(call)) {
817 		rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
818 					  0, -ENETRESET);
819 		return;
820 	}
821 
822 	/* If we get an OUT_OF_SEQUENCE ACK from the server, that can also
823 	 * indicate a change of address.  However, we can retransmit the call
824 	 * if we still have it buffered to the beginning.
825 	 */
826 	if (unlikely(ack.reason == RXRPC_ACK_OUT_OF_SEQUENCE) &&
827 	    first_soft_ack == 1 &&
828 	    prev_pkt == 0 &&
829 	    call->acks_hard_ack == 0 &&
830 	    rxrpc_is_client_call(call)) {
831 		rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
832 					  0, -ENETRESET);
833 		return;
834 	}
835 
836 	/* Discard any out-of-order or duplicate ACKs (outside lock). */
837 	if (!rxrpc_is_ack_valid(call, first_soft_ack, prev_pkt)) {
838 		trace_rxrpc_rx_discard_ack(call->debug_id, ack_serial,
839 					   first_soft_ack, call->acks_first_seq,
840 					   prev_pkt, call->acks_prev_seq);
841 		return;
842 	}
843 
844 	info.rxMTU = 0;
845 	ioffset = offset + nr_acks + 3;
846 	if (skb->len >= ioffset + sizeof(info) &&
847 	    skb_copy_bits(skb, ioffset, &info, sizeof(info)) < 0)
848 		return rxrpc_proto_abort("XAI", call, 0);
849 
850 	if (nr_acks > 0)
851 		skb_condense(skb);
852 
853 	call->acks_latest_ts = skb->tstamp;
854 	call->acks_first_seq = first_soft_ack;
855 	call->acks_prev_seq = prev_pkt;
856 
857 	switch (ack.reason) {
858 	case RXRPC_ACK_PING:
859 		break;
860 	default:
861 		if (after(acked_serial, call->acks_highest_serial))
862 			call->acks_highest_serial = acked_serial;
863 		break;
864 	}
865 
866 	/* Parse rwind and mtu sizes if provided. */
867 	if (info.rxMTU)
868 		rxrpc_input_ackinfo(call, skb, &info);
869 
870 	if (first_soft_ack == 0)
871 		return rxrpc_proto_abort("AK0", call, 0);
872 
873 	/* Ignore ACKs unless we are or have just been transmitting. */
874 	switch (READ_ONCE(call->state)) {
875 	case RXRPC_CALL_CLIENT_SEND_REQUEST:
876 	case RXRPC_CALL_CLIENT_AWAIT_REPLY:
877 	case RXRPC_CALL_SERVER_SEND_REPLY:
878 	case RXRPC_CALL_SERVER_AWAIT_ACK:
879 		break;
880 	default:
881 		return;
882 	}
883 
884 	if (before(hard_ack, call->acks_hard_ack) ||
885 	    after(hard_ack, call->tx_top))
886 		return rxrpc_proto_abort("AKW", call, 0);
887 	if (nr_acks > call->tx_top - hard_ack)
888 		return rxrpc_proto_abort("AKN", call, 0);
889 
890 	if (after(hard_ack, call->acks_hard_ack)) {
891 		if (rxrpc_rotate_tx_window(call, hard_ack, &summary)) {
892 			rxrpc_end_tx_phase(call, false, "ETA");
893 			return;
894 		}
895 	}
896 
897 	if (nr_acks > 0) {
898 		if (offset > (int)skb->len - nr_acks)
899 			return rxrpc_proto_abort("XSA", call, 0);
900 		rxrpc_input_soft_acks(call, skb->data + offset, first_soft_ack,
901 				      nr_acks, &summary);
902 	}
903 
904 	if (test_bit(RXRPC_CALL_TX_LAST, &call->flags) &&
905 	    summary.nr_acks == call->tx_top - hard_ack &&
906 	    rxrpc_is_client_call(call))
907 		rxrpc_propose_ping(call, ack_serial,
908 				   rxrpc_propose_ack_ping_for_lost_reply);
909 
910 	rxrpc_congestion_management(call, skb, &summary, acked_serial);
911 }
912 
913 /*
914  * Process an ACKALL packet.
915  */
916 static void rxrpc_input_ackall(struct rxrpc_call *call, struct sk_buff *skb)
917 {
918 	struct rxrpc_ack_summary summary = { 0 };
919 
920 	if (rxrpc_rotate_tx_window(call, call->tx_top, &summary))
921 		rxrpc_end_tx_phase(call, false, "ETL");
922 }
923 
924 /*
925  * Process an ABORT packet directed at a call.
926  */
927 static void rxrpc_input_abort(struct rxrpc_call *call, struct sk_buff *skb)
928 {
929 	struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
930 
931 	trace_rxrpc_rx_abort(call, sp->hdr.serial, skb->priority);
932 
933 	rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
934 				  skb->priority, -ECONNABORTED);
935 }
936 
937 /*
938  * Process an incoming call packet.
939  */
940 void rxrpc_input_call_packet(struct rxrpc_call *call, struct sk_buff *skb)
941 {
942 	struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
943 	unsigned long timo;
944 
945 	_enter("%p,%p", call, skb);
946 
947 	if (sp->hdr.serviceId != call->dest_srx.srx_service)
948 		call->dest_srx.srx_service = sp->hdr.serviceId;
949 	if ((int)sp->hdr.serial - (int)call->rx_serial > 0)
950 		call->rx_serial = sp->hdr.serial;
951 	if (!test_bit(RXRPC_CALL_RX_HEARD, &call->flags))
952 		set_bit(RXRPC_CALL_RX_HEARD, &call->flags);
953 
954 	timo = READ_ONCE(call->next_rx_timo);
955 	if (timo) {
956 		unsigned long now = jiffies, expect_rx_by;
957 
958 		expect_rx_by = now + timo;
959 		WRITE_ONCE(call->expect_rx_by, expect_rx_by);
960 		rxrpc_reduce_call_timer(call, expect_rx_by, now,
961 					rxrpc_timer_set_for_normal);
962 	}
963 
964 	switch (sp->hdr.type) {
965 	case RXRPC_PACKET_TYPE_DATA:
966 		rxrpc_input_data(call, skb);
967 		break;
968 
969 	case RXRPC_PACKET_TYPE_ACK:
970 		rxrpc_input_ack(call, skb);
971 		break;
972 
973 	case RXRPC_PACKET_TYPE_BUSY:
974 		/* Just ignore BUSY packets from the server; the retry and
975 		 * lifespan timers will take care of business.  BUSY packets
976 		 * from the client don't make sense.
977 		 */
978 		break;
979 
980 	case RXRPC_PACKET_TYPE_ABORT:
981 		rxrpc_input_abort(call, skb);
982 		break;
983 
984 	case RXRPC_PACKET_TYPE_ACKALL:
985 		rxrpc_input_ackall(call, skb);
986 		break;
987 
988 	default:
989 		break;
990 	}
991 }
992 
993 /*
994  * Handle a new service call on a channel implicitly completing the preceding
995  * call on that channel.  This does not apply to client conns.
996  *
997  * TODO: If callNumber > call_id + 1, renegotiate security.
998  */
999 void rxrpc_implicit_end_call(struct rxrpc_call *call, struct sk_buff *skb)
1000 {
1001 	struct rxrpc_connection *conn = call->conn;
1002 
1003 	switch (READ_ONCE(call->state)) {
1004 	case RXRPC_CALL_SERVER_AWAIT_ACK:
1005 		rxrpc_call_completed(call);
1006 		fallthrough;
1007 	case RXRPC_CALL_COMPLETE:
1008 		break;
1009 	default:
1010 		if (rxrpc_abort_call("IMP", call, 0, RX_CALL_DEAD, -ESHUTDOWN))
1011 			rxrpc_send_abort_packet(call);
1012 		trace_rxrpc_improper_term(call);
1013 		break;
1014 	}
1015 
1016 	rxrpc_input_call_event(call, skb);
1017 
1018 	spin_lock(&conn->bundle->channel_lock);
1019 	__rxrpc_disconnect_call(conn, call);
1020 	spin_unlock(&conn->bundle->channel_lock);
1021 }
1022