Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[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         read_lock_bh(&call->state_lock);
35
36         if (call->state < RXRPC_CALL_COMPLETE) {
37                 t = call->expire_at;
38                 if (!ktime_after(t, now))
39                         goto out;
40
41                 if (!ktime_after(call->resend_at, now)) {
42                         call->resend_at = call->expire_at;
43                         if (!test_and_set_bit(RXRPC_CALL_EV_RESEND, &call->events))
44                                 queue = true;
45                 } else if (ktime_before(call->resend_at, t)) {
46                         t = call->resend_at;
47                 }
48
49                 if (!ktime_after(call->ack_at, now)) {
50                         call->ack_at = call->expire_at;
51                         if (!test_and_set_bit(RXRPC_CALL_EV_ACK, &call->events))
52                                 queue = true;
53                 } else if (ktime_before(call->ack_at, t)) {
54                         t = call->ack_at;
55                 }
56
57                 t_j = nsecs_to_jiffies(ktime_to_ns(ktime_sub(t, now)));
58                 t_j += jiffies;
59
60                 /* We have to make sure that the calculated jiffies value falls
61                  * at or after the nsec value, or we may loop ceaselessly
62                  * because the timer times out, but we haven't reached the nsec
63                  * timeout yet.
64                  */
65                 t_j++;
66
67                 if (call->timer.expires != t_j || !timer_pending(&call->timer)) {
68                         mod_timer(&call->timer, t_j);
69                         trace_rxrpc_timer(call, why, now, now_j);
70                 }
71
72                 if (queue)
73                         rxrpc_queue_call(call);
74         }
75
76 out:
77         read_unlock_bh(&call->state_lock);
78 }
79
80 /*
81  * propose an ACK be sent
82  */
83 static void __rxrpc_propose_ACK(struct rxrpc_call *call, u8 ack_reason,
84                                 u16 skew, u32 serial, bool immediate,
85                                 bool background,
86                                 enum rxrpc_propose_ack_trace why)
87 {
88         enum rxrpc_propose_ack_outcome outcome = rxrpc_propose_ack_use;
89         unsigned int expiry = rxrpc_soft_ack_delay;
90         ktime_t now, ack_at;
91         s8 prior = rxrpc_ack_priority[ack_reason];
92
93         /* Update DELAY, IDLE, REQUESTED and PING_RESPONSE ACK serial
94          * numbers, but we don't alter the timeout.
95          */
96         _debug("prior %u %u vs %u %u",
97                ack_reason, prior,
98                call->ackr_reason, rxrpc_ack_priority[call->ackr_reason]);
99         if (ack_reason == call->ackr_reason) {
100                 if (RXRPC_ACK_UPDATEABLE & (1 << ack_reason)) {
101                         outcome = rxrpc_propose_ack_update;
102                         call->ackr_serial = serial;
103                         call->ackr_skew = skew;
104                 }
105                 if (!immediate)
106                         goto trace;
107         } else if (prior > rxrpc_ack_priority[call->ackr_reason]) {
108                 call->ackr_reason = ack_reason;
109                 call->ackr_serial = serial;
110                 call->ackr_skew = skew;
111         } else {
112                 outcome = rxrpc_propose_ack_subsume;
113         }
114
115         switch (ack_reason) {
116         case RXRPC_ACK_REQUESTED:
117                 if (rxrpc_requested_ack_delay < expiry)
118                         expiry = rxrpc_requested_ack_delay;
119                 if (serial == 1)
120                         immediate = false;
121                 break;
122
123         case RXRPC_ACK_DELAY:
124                 if (rxrpc_soft_ack_delay < expiry)
125                         expiry = rxrpc_soft_ack_delay;
126                 break;
127
128         case RXRPC_ACK_PING:
129         case RXRPC_ACK_IDLE:
130                 if (rxrpc_idle_ack_delay < expiry)
131                         expiry = rxrpc_idle_ack_delay;
132                 break;
133
134         default:
135                 immediate = true;
136                 break;
137         }
138
139         if (test_bit(RXRPC_CALL_EV_ACK, &call->events)) {
140                 _debug("already scheduled");
141         } else if (immediate || expiry == 0) {
142                 _debug("immediate ACK %lx", call->events);
143                 if (!test_and_set_bit(RXRPC_CALL_EV_ACK, &call->events) &&
144                     background)
145                         rxrpc_queue_call(call);
146         } else {
147                 now = ktime_get_real();
148                 ack_at = ktime_add_ms(now, expiry);
149                 if (ktime_before(ack_at, call->ack_at)) {
150                         call->ack_at = ack_at;
151                         rxrpc_set_timer(call, rxrpc_timer_set_for_ack, now);
152                 }
153         }
154
155 trace:
156         trace_rxrpc_propose_ack(call, why, ack_reason, serial, immediate,
157                                 background, outcome);
158 }
159
160 /*
161  * propose an ACK be sent, locking the call structure
162  */
163 void rxrpc_propose_ACK(struct rxrpc_call *call, u8 ack_reason,
164                        u16 skew, u32 serial, bool immediate, bool background,
165                        enum rxrpc_propose_ack_trace why)
166 {
167         spin_lock_bh(&call->lock);
168         __rxrpc_propose_ACK(call, ack_reason, skew, serial,
169                             immediate, background, why);
170         spin_unlock_bh(&call->lock);
171 }
172
173 /*
174  * Handle congestion being detected by the retransmit timeout.
175  */
176 static void rxrpc_congestion_timeout(struct rxrpc_call *call)
177 {
178         set_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags);
179 }
180
181 /*
182  * Perform retransmission of NAK'd and unack'd packets.
183  */
184 static void rxrpc_resend(struct rxrpc_call *call, ktime_t now)
185 {
186         struct rxrpc_skb_priv *sp;
187         struct sk_buff *skb;
188         rxrpc_seq_t cursor, seq, top;
189         ktime_t max_age, oldest, ack_ts;
190         int ix;
191         u8 annotation, anno_type, retrans = 0, unacked = 0;
192
193         _enter("{%d,%d}", call->tx_hard_ack, call->tx_top);
194
195         max_age = ktime_sub_ms(now, rxrpc_resend_timeout);
196
197         spin_lock_bh(&call->lock);
198
199         cursor = call->tx_hard_ack;
200         top = call->tx_top;
201         ASSERT(before_eq(cursor, top));
202         if (cursor == top)
203                 goto out_unlock;
204
205         /* Scan the packet list without dropping the lock and decide which of
206          * the packets in the Tx buffer we're going to resend and what the new
207          * resend timeout will be.
208          */
209         oldest = now;
210         for (seq = cursor + 1; before_eq(seq, top); seq++) {
211                 ix = seq & RXRPC_RXTX_BUFF_MASK;
212                 annotation = call->rxtx_annotations[ix];
213                 anno_type = annotation & RXRPC_TX_ANNO_MASK;
214                 annotation &= ~RXRPC_TX_ANNO_MASK;
215                 if (anno_type == RXRPC_TX_ANNO_ACK)
216                         continue;
217
218                 skb = call->rxtx_buffer[ix];
219                 rxrpc_see_skb(skb, rxrpc_skb_tx_seen);
220                 sp = rxrpc_skb(skb);
221
222                 if (anno_type == RXRPC_TX_ANNO_UNACK) {
223                         if (ktime_after(skb->tstamp, max_age)) {
224                                 if (ktime_before(skb->tstamp, oldest))
225                                         oldest = skb->tstamp;
226                                 continue;
227                         }
228                         if (!(annotation & RXRPC_TX_ANNO_RESENT))
229                                 unacked++;
230                 }
231
232                 /* Okay, we need to retransmit a packet. */
233                 call->rxtx_annotations[ix] = RXRPC_TX_ANNO_RETRANS | annotation;
234                 retrans++;
235                 trace_rxrpc_retransmit(call, seq, annotation | anno_type,
236                                        ktime_to_ns(ktime_sub(skb->tstamp, max_age)));
237         }
238
239         call->resend_at = ktime_add_ms(oldest, rxrpc_resend_timeout);
240
241         if (unacked)
242                 rxrpc_congestion_timeout(call);
243
244         /* If there was nothing that needed retransmission then it's likely
245          * that an ACK got lost somewhere.  Send a ping to find out instead of
246          * retransmitting data.
247          */
248         if (!retrans) {
249                 rxrpc_set_timer(call, rxrpc_timer_set_for_resend, now);
250                 spin_unlock_bh(&call->lock);
251                 ack_ts = ktime_sub(now, call->acks_latest_ts);
252                 if (ktime_to_ns(ack_ts) < call->peer->rtt)
253                         goto out;
254                 rxrpc_propose_ACK(call, RXRPC_ACK_PING, 0, 0, true, false,
255                                   rxrpc_propose_ack_ping_for_lost_ack);
256                 rxrpc_send_call_packet(call, RXRPC_PACKET_TYPE_ACK);
257                 goto out;
258         }
259
260         /* Now go through the Tx window and perform the retransmissions.  We
261          * have to drop the lock for each send.  If an ACK comes in whilst the
262          * lock is dropped, it may clear some of the retransmission markers for
263          * packets that it soft-ACKs.
264          */
265         for (seq = cursor + 1; before_eq(seq, top); seq++) {
266                 ix = seq & RXRPC_RXTX_BUFF_MASK;
267                 annotation = call->rxtx_annotations[ix];
268                 anno_type = annotation & RXRPC_TX_ANNO_MASK;
269                 if (anno_type != RXRPC_TX_ANNO_RETRANS)
270                         continue;
271
272                 skb = call->rxtx_buffer[ix];
273                 rxrpc_get_skb(skb, rxrpc_skb_tx_got);
274                 spin_unlock_bh(&call->lock);
275
276                 if (rxrpc_send_data_packet(call, skb, true) < 0) {
277                         rxrpc_free_skb(skb, rxrpc_skb_tx_freed);
278                         return;
279                 }
280
281                 if (rxrpc_is_client_call(call))
282                         rxrpc_expose_client_call(call);
283
284                 rxrpc_free_skb(skb, rxrpc_skb_tx_freed);
285                 spin_lock_bh(&call->lock);
286
287                 /* We need to clear the retransmit state, but there are two
288                  * things we need to be aware of: A new ACK/NAK might have been
289                  * received and the packet might have been hard-ACK'd (in which
290                  * case it will no longer be in the buffer).
291                  */
292                 if (after(seq, call->tx_hard_ack)) {
293                         annotation = call->rxtx_annotations[ix];
294                         anno_type = annotation & RXRPC_TX_ANNO_MASK;
295                         if (anno_type == RXRPC_TX_ANNO_RETRANS ||
296                             anno_type == RXRPC_TX_ANNO_NAK) {
297                                 annotation &= ~RXRPC_TX_ANNO_MASK;
298                                 annotation |= RXRPC_TX_ANNO_UNACK;
299                         }
300                         annotation |= RXRPC_TX_ANNO_RESENT;
301                         call->rxtx_annotations[ix] = annotation;
302                 }
303
304                 if (after(call->tx_hard_ack, seq))
305                         seq = call->tx_hard_ack;
306         }
307
308 out_unlock:
309         spin_unlock_bh(&call->lock);
310 out:
311         _leave("");
312 }
313
314 /*
315  * Handle retransmission and deferred ACK/abort generation.
316  */
317 void rxrpc_process_call(struct work_struct *work)
318 {
319         struct rxrpc_call *call =
320                 container_of(work, struct rxrpc_call, processor);
321         ktime_t now;
322
323         rxrpc_see_call(call);
324
325         //printk("\n--------------------\n");
326         _enter("{%d,%s,%lx}",
327                call->debug_id, rxrpc_call_states[call->state], call->events);
328
329 recheck_state:
330         if (test_and_clear_bit(RXRPC_CALL_EV_ABORT, &call->events)) {
331                 rxrpc_send_call_packet(call, RXRPC_PACKET_TYPE_ABORT);
332                 goto recheck_state;
333         }
334
335         if (call->state == RXRPC_CALL_COMPLETE) {
336                 del_timer_sync(&call->timer);
337                 goto out_put;
338         }
339
340         now = ktime_get_real();
341         if (ktime_before(call->expire_at, now)) {
342                 rxrpc_abort_call("EXP", call, 0, RX_CALL_TIMEOUT, ETIME);
343                 set_bit(RXRPC_CALL_EV_ABORT, &call->events);
344                 goto recheck_state;
345         }
346
347         if (test_and_clear_bit(RXRPC_CALL_EV_ACK, &call->events)) {
348                 call->ack_at = call->expire_at;
349                 if (call->ackr_reason) {
350                         rxrpc_send_call_packet(call, RXRPC_PACKET_TYPE_ACK);
351                         goto recheck_state;
352                 }
353         }
354
355         if (test_and_clear_bit(RXRPC_CALL_EV_RESEND, &call->events)) {
356                 rxrpc_resend(call, now);
357                 goto recheck_state;
358         }
359
360         rxrpc_set_timer(call, rxrpc_timer_set_for_resend, now);
361
362         /* other events may have been raised since we started checking */
363         if (call->events && call->state < RXRPC_CALL_COMPLETE) {
364                 __rxrpc_queue_call(call);
365                 goto out;
366         }
367
368 out_put:
369         rxrpc_put_call(call, rxrpc_call_put);
370 out:
371         _leave("");
372 }