Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
[cascardo/linux.git] / net / rxrpc / call_event.c
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  * Set the timer
26  */
27 void __rxrpc_set_timer(struct rxrpc_call *call, enum rxrpc_timer_trace why,
28                        ktime_t now)
29 {
30         unsigned long t_j, now_j = jiffies;
31         ktime_t t;
32         bool queue = false;
33
34         if (call->state < RXRPC_CALL_COMPLETE) {
35                 t = call->expire_at;
36                 if (!ktime_after(t, now)) {
37                         trace_rxrpc_timer(call, why, now, now_j);
38                         queue = true;
39                         goto out;
40                 }
41
42                 if (!ktime_after(call->resend_at, now)) {
43                         call->resend_at = call->expire_at;
44                         if (!test_and_set_bit(RXRPC_CALL_EV_RESEND, &call->events))
45                                 queue = true;
46                 } else if (ktime_before(call->resend_at, t)) {
47                         t = call->resend_at;
48                 }
49
50                 if (!ktime_after(call->ack_at, now)) {
51                         call->ack_at = call->expire_at;
52                         if (!test_and_set_bit(RXRPC_CALL_EV_ACK, &call->events))
53                                 queue = true;
54                 } else if (ktime_before(call->ack_at, t)) {
55                         t = call->ack_at;
56                 }
57
58                 if (!ktime_after(call->ping_at, now)) {
59                         call->ping_at = call->expire_at;
60                         if (!test_and_set_bit(RXRPC_CALL_EV_PING, &call->events))
61                                 queue = true;
62                 } else if (ktime_before(call->ping_at, t)) {
63                         t = call->ping_at;
64                 }
65
66                 t_j = nsecs_to_jiffies(ktime_to_ns(ktime_sub(t, now)));
67                 t_j += jiffies;
68
69                 /* We have to make sure that the calculated jiffies value falls
70                  * at or after the nsec value, or we may loop ceaselessly
71                  * because the timer times out, but we haven't reached the nsec
72                  * timeout yet.
73                  */
74                 t_j++;
75
76                 if (call->timer.expires != t_j || !timer_pending(&call->timer)) {
77                         mod_timer(&call->timer, t_j);
78                         trace_rxrpc_timer(call, why, now, now_j);
79                 }
80         }
81
82 out:
83         if (queue)
84                 rxrpc_queue_call(call);
85 }
86
87 /*
88  * Set the timer
89  */
90 void rxrpc_set_timer(struct rxrpc_call *call, enum rxrpc_timer_trace why,
91                      ktime_t now)
92 {
93         read_lock_bh(&call->state_lock);
94         __rxrpc_set_timer(call, why, now);
95         read_unlock_bh(&call->state_lock);
96 }
97
98 /*
99  * Propose a PING ACK be sent.
100  */
101 static void rxrpc_propose_ping(struct rxrpc_call *call,
102                                bool immediate, bool background)
103 {
104         if (immediate) {
105                 if (background &&
106                     !test_and_set_bit(RXRPC_CALL_EV_PING, &call->events))
107                         rxrpc_queue_call(call);
108         } else {
109                 ktime_t now = ktime_get_real();
110                 ktime_t ping_at = ktime_add_ms(now, rxrpc_idle_ack_delay);
111
112                 if (ktime_before(ping_at, call->ping_at)) {
113                         call->ping_at = ping_at;
114                         rxrpc_set_timer(call, rxrpc_timer_set_for_ping, now);
115                 }
116         }
117 }
118
119 /*
120  * propose an ACK be sent
121  */
122 static void __rxrpc_propose_ACK(struct rxrpc_call *call, u8 ack_reason,
123                                 u16 skew, u32 serial, bool immediate,
124                                 bool background,
125                                 enum rxrpc_propose_ack_trace why)
126 {
127         enum rxrpc_propose_ack_outcome outcome = rxrpc_propose_ack_use;
128         unsigned int expiry = rxrpc_soft_ack_delay;
129         ktime_t now, ack_at;
130         s8 prior = rxrpc_ack_priority[ack_reason];
131
132         /* Pings are handled specially because we don't want to accidentally
133          * lose a ping response by subsuming it into a ping.
134          */
135         if (ack_reason == RXRPC_ACK_PING) {
136                 rxrpc_propose_ping(call, immediate, background);
137                 goto trace;
138         }
139
140         /* Update DELAY, IDLE, REQUESTED and PING_RESPONSE ACK serial
141          * numbers, but we don't alter the timeout.
142          */
143         _debug("prior %u %u vs %u %u",
144                ack_reason, prior,
145                call->ackr_reason, rxrpc_ack_priority[call->ackr_reason]);
146         if (ack_reason == call->ackr_reason) {
147                 if (RXRPC_ACK_UPDATEABLE & (1 << ack_reason)) {
148                         outcome = rxrpc_propose_ack_update;
149                         call->ackr_serial = serial;
150                         call->ackr_skew = skew;
151                 }
152                 if (!immediate)
153                         goto trace;
154         } else if (prior > rxrpc_ack_priority[call->ackr_reason]) {
155                 call->ackr_reason = ack_reason;
156                 call->ackr_serial = serial;
157                 call->ackr_skew = skew;
158         } else {
159                 outcome = rxrpc_propose_ack_subsume;
160         }
161
162         switch (ack_reason) {
163         case RXRPC_ACK_REQUESTED:
164                 if (rxrpc_requested_ack_delay < expiry)
165                         expiry = rxrpc_requested_ack_delay;
166                 if (serial == 1)
167                         immediate = false;
168                 break;
169
170         case RXRPC_ACK_DELAY:
171                 if (rxrpc_soft_ack_delay < expiry)
172                         expiry = rxrpc_soft_ack_delay;
173                 break;
174
175         case RXRPC_ACK_IDLE:
176                 if (rxrpc_idle_ack_delay < expiry)
177                         expiry = rxrpc_idle_ack_delay;
178                 break;
179
180         default:
181                 immediate = true;
182                 break;
183         }
184
185         if (test_bit(RXRPC_CALL_EV_ACK, &call->events)) {
186                 _debug("already scheduled");
187         } else if (immediate || expiry == 0) {
188                 _debug("immediate ACK %lx", call->events);
189                 if (!test_and_set_bit(RXRPC_CALL_EV_ACK, &call->events) &&
190                     background)
191                         rxrpc_queue_call(call);
192         } else {
193                 now = ktime_get_real();
194                 ack_at = ktime_add_ms(now, expiry);
195                 if (ktime_before(ack_at, call->ack_at)) {
196                         call->ack_at = ack_at;
197                         rxrpc_set_timer(call, rxrpc_timer_set_for_ack, now);
198                 }
199         }
200
201 trace:
202         trace_rxrpc_propose_ack(call, why, ack_reason, serial, immediate,
203                                 background, outcome);
204 }
205
206 /*
207  * propose an ACK be sent, locking the call structure
208  */
209 void rxrpc_propose_ACK(struct rxrpc_call *call, u8 ack_reason,
210                        u16 skew, u32 serial, bool immediate, bool background,
211                        enum rxrpc_propose_ack_trace why)
212 {
213         spin_lock_bh(&call->lock);
214         __rxrpc_propose_ACK(call, ack_reason, skew, serial,
215                             immediate, background, why);
216         spin_unlock_bh(&call->lock);
217 }
218
219 /*
220  * Handle congestion being detected by the retransmit timeout.
221  */
222 static void rxrpc_congestion_timeout(struct rxrpc_call *call)
223 {
224         set_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags);
225 }
226
227 /*
228  * Perform retransmission of NAK'd and unack'd packets.
229  */
230 static void rxrpc_resend(struct rxrpc_call *call, ktime_t now)
231 {
232         struct rxrpc_skb_priv *sp;
233         struct sk_buff *skb;
234         rxrpc_seq_t cursor, seq, top;
235         ktime_t max_age, oldest, ack_ts;
236         int ix;
237         u8 annotation, anno_type, retrans = 0, unacked = 0;
238
239         _enter("{%d,%d}", call->tx_hard_ack, call->tx_top);
240
241         max_age = ktime_sub_ms(now, rxrpc_resend_timeout);
242
243         spin_lock_bh(&call->lock);
244
245         cursor = call->tx_hard_ack;
246         top = call->tx_top;
247         ASSERT(before_eq(cursor, top));
248         if (cursor == top)
249                 goto out_unlock;
250
251         /* Scan the packet list without dropping the lock and decide which of
252          * the packets in the Tx buffer we're going to resend and what the new
253          * resend timeout will be.
254          */
255         oldest = now;
256         for (seq = cursor + 1; before_eq(seq, top); seq++) {
257                 ix = seq & RXRPC_RXTX_BUFF_MASK;
258                 annotation = call->rxtx_annotations[ix];
259                 anno_type = annotation & RXRPC_TX_ANNO_MASK;
260                 annotation &= ~RXRPC_TX_ANNO_MASK;
261                 if (anno_type == RXRPC_TX_ANNO_ACK)
262                         continue;
263
264                 skb = call->rxtx_buffer[ix];
265                 rxrpc_see_skb(skb, rxrpc_skb_tx_seen);
266                 sp = rxrpc_skb(skb);
267
268                 if (anno_type == RXRPC_TX_ANNO_UNACK) {
269                         if (ktime_after(skb->tstamp, max_age)) {
270                                 if (ktime_before(skb->tstamp, oldest))
271                                         oldest = skb->tstamp;
272                                 continue;
273                         }
274                         if (!(annotation & RXRPC_TX_ANNO_RESENT))
275                                 unacked++;
276                 }
277
278                 /* Okay, we need to retransmit a packet. */
279                 call->rxtx_annotations[ix] = RXRPC_TX_ANNO_RETRANS | annotation;
280                 retrans++;
281                 trace_rxrpc_retransmit(call, seq, annotation | anno_type,
282                                        ktime_to_ns(ktime_sub(skb->tstamp, max_age)));
283         }
284
285         call->resend_at = ktime_add_ms(oldest, rxrpc_resend_timeout);
286
287         if (unacked)
288                 rxrpc_congestion_timeout(call);
289
290         /* If there was nothing that needed retransmission then it's likely
291          * that an ACK got lost somewhere.  Send a ping to find out instead of
292          * retransmitting data.
293          */
294         if (!retrans) {
295                 rxrpc_set_timer(call, rxrpc_timer_set_for_resend, now);
296                 spin_unlock_bh(&call->lock);
297                 ack_ts = ktime_sub(now, call->acks_latest_ts);
298                 if (ktime_to_ns(ack_ts) < call->peer->rtt)
299                         goto out;
300                 rxrpc_propose_ACK(call, RXRPC_ACK_PING, 0, 0, true, false,
301                                   rxrpc_propose_ack_ping_for_lost_ack);
302                 rxrpc_send_ack_packet(call, true);
303                 goto out;
304         }
305
306         /* Now go through the Tx window and perform the retransmissions.  We
307          * have to drop the lock for each send.  If an ACK comes in whilst the
308          * lock is dropped, it may clear some of the retransmission markers for
309          * packets that it soft-ACKs.
310          */
311         for (seq = cursor + 1; before_eq(seq, top); seq++) {
312                 ix = seq & RXRPC_RXTX_BUFF_MASK;
313                 annotation = call->rxtx_annotations[ix];
314                 anno_type = annotation & RXRPC_TX_ANNO_MASK;
315                 if (anno_type != RXRPC_TX_ANNO_RETRANS)
316                         continue;
317
318                 skb = call->rxtx_buffer[ix];
319                 rxrpc_get_skb(skb, rxrpc_skb_tx_got);
320                 spin_unlock_bh(&call->lock);
321
322                 if (rxrpc_send_data_packet(call, skb, true) < 0) {
323                         rxrpc_free_skb(skb, rxrpc_skb_tx_freed);
324                         return;
325                 }
326
327                 if (rxrpc_is_client_call(call))
328                         rxrpc_expose_client_call(call);
329
330                 rxrpc_free_skb(skb, rxrpc_skb_tx_freed);
331                 spin_lock_bh(&call->lock);
332
333                 /* We need to clear the retransmit state, but there are two
334                  * things we need to be aware of: A new ACK/NAK might have been
335                  * received and the packet might have been hard-ACK'd (in which
336                  * case it will no longer be in the buffer).
337                  */
338                 if (after(seq, call->tx_hard_ack)) {
339                         annotation = call->rxtx_annotations[ix];
340                         anno_type = annotation & RXRPC_TX_ANNO_MASK;
341                         if (anno_type == RXRPC_TX_ANNO_RETRANS ||
342                             anno_type == RXRPC_TX_ANNO_NAK) {
343                                 annotation &= ~RXRPC_TX_ANNO_MASK;
344                                 annotation |= RXRPC_TX_ANNO_UNACK;
345                         }
346                         annotation |= RXRPC_TX_ANNO_RESENT;
347                         call->rxtx_annotations[ix] = annotation;
348                 }
349
350                 if (after(call->tx_hard_ack, seq))
351                         seq = call->tx_hard_ack;
352         }
353
354 out_unlock:
355         spin_unlock_bh(&call->lock);
356 out:
357         _leave("");
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         ktime_t now;
368
369         rxrpc_see_call(call);
370
371         //printk("\n--------------------\n");
372         _enter("{%d,%s,%lx}",
373                call->debug_id, rxrpc_call_states[call->state], call->events);
374
375 recheck_state:
376         if (test_and_clear_bit(RXRPC_CALL_EV_ABORT, &call->events)) {
377                 rxrpc_send_abort_packet(call);
378                 goto recheck_state;
379         }
380
381         if (call->state == RXRPC_CALL_COMPLETE) {
382                 del_timer_sync(&call->timer);
383                 rxrpc_notify_socket(call);
384                 goto out_put;
385         }
386
387         now = ktime_get_real();
388         if (ktime_before(call->expire_at, now)) {
389                 rxrpc_abort_call("EXP", call, 0, RX_CALL_TIMEOUT, ETIME);
390                 set_bit(RXRPC_CALL_EV_ABORT, &call->events);
391                 goto recheck_state;
392         }
393
394         if (test_and_clear_bit(RXRPC_CALL_EV_ACK, &call->events)) {
395                 if (call->ackr_reason) {
396                         rxrpc_send_ack_packet(call, false);
397                         goto recheck_state;
398                 }
399         }
400
401         if (test_and_clear_bit(RXRPC_CALL_EV_PING, &call->events)) {
402                 rxrpc_send_ack_packet(call, true);
403                 goto recheck_state;
404         }
405
406         if (test_and_clear_bit(RXRPC_CALL_EV_RESEND, &call->events)) {
407                 rxrpc_resend(call, now);
408                 goto recheck_state;
409         }
410
411         rxrpc_set_timer(call, rxrpc_timer_set_for_resend, now);
412
413         /* other events may have been raised since we started checking */
414         if (call->events && call->state < RXRPC_CALL_COMPLETE) {
415                 __rxrpc_queue_call(call);
416                 goto out;
417         }
418
419 out_put:
420         rxrpc_put_call(call, rxrpc_call_put);
421 out:
422         _leave("");
423 }