Merge master.kernel.org:/pub/scm/linux/kernel/git/sridhar/lksctp-2.6
[linux-2.6] / net / dccp / output.c
1 /*
2  *  net/dccp/output.c
3  * 
4  *  An implementation of the DCCP protocol
5  *  Arnaldo Carvalho de Melo <acme@conectiva.com.br>
6  *
7  *      This program is free software; you can redistribute it and/or
8  *      modify it under the terms of the GNU General Public License
9  *      as published by the Free Software Foundation; either version
10  *      2 of the License, or (at your option) any later version.
11  */
12
13 #include <linux/config.h>
14 #include <linux/dccp.h>
15 #include <linux/skbuff.h>
16
17 #include <net/sock.h>
18
19 #include "ackvec.h"
20 #include "ccid.h"
21 #include "dccp.h"
22
23 static inline void dccp_event_ack_sent(struct sock *sk)
24 {
25         inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
26 }
27
28 /*
29  * All SKB's seen here are completely headerless. It is our
30  * job to build the DCCP header, and pass the packet down to
31  * IP so it can do the same plus pass the packet off to the
32  * device.
33  */
34 int dccp_transmit_skb(struct sock *sk, struct sk_buff *skb)
35 {
36         if (likely(skb != NULL)) {
37                 const struct inet_sock *inet = inet_sk(sk);
38                 struct dccp_sock *dp = dccp_sk(sk);
39                 struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
40                 struct dccp_hdr *dh;
41                 /* XXX For now we're using only 48 bits sequence numbers */
42                 const int dccp_header_size = sizeof(*dh) +
43                                              sizeof(struct dccp_hdr_ext) +
44                                           dccp_packet_hdr_len(dcb->dccpd_type);
45                 int err, set_ack = 1;
46                 u64 ackno = dp->dccps_gsr;
47
48                 dccp_inc_seqno(&dp->dccps_gss);
49
50                 switch (dcb->dccpd_type) {
51                 case DCCP_PKT_DATA:
52                         set_ack = 0;
53                         break;
54                 case DCCP_PKT_SYNC:
55                 case DCCP_PKT_SYNCACK:
56                         ackno = dcb->dccpd_seq;
57                         break;
58                 }
59
60                 dcb->dccpd_seq = dp->dccps_gss;
61                 dccp_insert_options(sk, skb);
62                 
63                 skb->h.raw = skb_push(skb, dccp_header_size);
64                 dh = dccp_hdr(skb);
65
66                 if (!skb->sk)
67                         skb_set_owner_w(skb, sk);
68
69                 /* Build DCCP header and checksum it. */
70                 memset(dh, 0, dccp_header_size);
71                 dh->dccph_type  = dcb->dccpd_type;
72                 dh->dccph_sport = inet->sport;
73                 dh->dccph_dport = inet->dport;
74                 dh->dccph_doff  = (dccp_header_size + dcb->dccpd_opt_len) / 4;
75                 dh->dccph_ccval = dcb->dccpd_ccval;
76                 /* XXX For now we're using only 48 bits sequence numbers */
77                 dh->dccph_x     = 1;
78
79                 dp->dccps_awh = dp->dccps_gss;
80                 dccp_hdr_set_seq(dh, dp->dccps_gss);
81                 if (set_ack)
82                         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), ackno);
83
84                 switch (dcb->dccpd_type) {
85                 case DCCP_PKT_REQUEST:
86                         dccp_hdr_request(skb)->dccph_req_service =
87                                                         dp->dccps_service;
88                         break;
89                 case DCCP_PKT_RESET:
90                         dccp_hdr_reset(skb)->dccph_reset_code =
91                                                         dcb->dccpd_reset_code;
92                         break;
93                 }
94
95                 dh->dccph_checksum = dccp_v4_checksum(skb, inet->saddr,
96                                                       inet->daddr);
97
98                 if (set_ack)
99                         dccp_event_ack_sent(sk);
100
101                 DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
102
103                 memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
104                 err = ip_queue_xmit(skb, 0);
105                 if (err <= 0)
106                         return err;
107
108                 /* NET_XMIT_CN is special. It does not guarantee,
109                  * that this packet is lost. It tells that device
110                  * is about to start to drop packets or already
111                  * drops some packets of the same priority and
112                  * invokes us to send less aggressively.
113                  */
114                 return err == NET_XMIT_CN ? 0 : err;
115         }
116         return -ENOBUFS;
117 }
118
119 unsigned int dccp_sync_mss(struct sock *sk, u32 pmtu)
120 {
121         struct dccp_sock *dp = dccp_sk(sk);
122         int mss_now;
123
124         /*
125          * FIXME: we really should be using the af_specific thing to support
126          *        IPv6.
127          * mss_now = pmtu - tp->af_specific->net_header_len -
128          *           sizeof(struct dccp_hdr) - sizeof(struct dccp_hdr_ext);
129          */
130         mss_now = pmtu - sizeof(struct iphdr) - sizeof(struct dccp_hdr) -
131                   sizeof(struct dccp_hdr_ext);
132
133         /* Now subtract optional transport overhead */
134         mss_now -= dp->dccps_ext_header_len;
135
136         /*
137          * FIXME: this should come from the CCID infrastructure, where, say,
138          * TFRC will say it wants TIMESTAMPS, ELAPSED time, etc, for now lets
139          * put a rough estimate for NDP + TIMESTAMP + TIMESTAMP_ECHO + ELAPSED
140          * TIME + TFRC_OPT_LOSS_EVENT_RATE + TFRC_OPT_RECEIVE_RATE + padding to
141          * make it a multiple of 4
142          */
143
144         mss_now -= ((5 + 6 + 10 + 6 + 6 + 6 + 3) / 4) * 4;
145
146         /* And store cached results */
147         dp->dccps_pmtu_cookie = pmtu;
148         dp->dccps_mss_cache = mss_now;
149
150         return mss_now;
151 }
152
153 void dccp_write_space(struct sock *sk)
154 {
155         read_lock(&sk->sk_callback_lock);
156
157         if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
158                 wake_up_interruptible(sk->sk_sleep);
159         /* Should agree with poll, otherwise some programs break */
160         if (sock_writeable(sk))
161                 sk_wake_async(sk, 2, POLL_OUT);
162
163         read_unlock(&sk->sk_callback_lock);
164 }
165
166 /**
167  * dccp_wait_for_ccid - Wait for ccid to tell us we can send a packet
168  * @sk: socket to wait for
169  * @timeo: for how long
170  */
171 static int dccp_wait_for_ccid(struct sock *sk, struct sk_buff *skb,
172                               long *timeo)
173 {
174         struct dccp_sock *dp = dccp_sk(sk);
175         DEFINE_WAIT(wait);
176         long delay;
177         int rc;
178
179         while (1) {
180                 prepare_to_wait(sk->sk_sleep, &wait, TASK_INTERRUPTIBLE);
181
182                 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
183                         goto do_error;
184                 if (!*timeo)
185                         goto do_nonblock;
186                 if (signal_pending(current))
187                         goto do_interrupted;
188
189                 rc = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
190                                             skb->len);
191                 if (rc <= 0)
192                         break;
193                 delay = msecs_to_jiffies(rc);
194                 if (delay > *timeo || delay < 0)
195                         goto do_nonblock;
196
197                 sk->sk_write_pending++;
198                 release_sock(sk);
199                 *timeo -= schedule_timeout(delay);
200                 lock_sock(sk);
201                 sk->sk_write_pending--;
202         }
203 out:
204         finish_wait(sk->sk_sleep, &wait);
205         return rc;
206
207 do_error:
208         rc = -EPIPE;
209         goto out;
210 do_nonblock:
211         rc = -EAGAIN;
212         goto out;
213 do_interrupted:
214         rc = sock_intr_errno(*timeo);
215         goto out;
216 }
217
218 int dccp_write_xmit(struct sock *sk, struct sk_buff *skb, long *timeo)
219 {
220         const struct dccp_sock *dp = dccp_sk(sk);
221         int err = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
222                                          skb->len);
223
224         if (err > 0)
225                 err = dccp_wait_for_ccid(sk, skb, timeo);
226
227         if (err == 0) {
228                 struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
229                 const int len = skb->len;
230
231                 if (sk->sk_state == DCCP_PARTOPEN) {
232                         /* See 8.1.5.  Handshake Completion */
233                         inet_csk_schedule_ack(sk);
234                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
235                                                   inet_csk(sk)->icsk_rto,
236                                                   DCCP_RTO_MAX);
237                         dcb->dccpd_type = DCCP_PKT_DATAACK;
238                 } else if (dccp_ack_pending(sk))
239                         dcb->dccpd_type = DCCP_PKT_DATAACK;
240                 else
241                         dcb->dccpd_type = DCCP_PKT_DATA;
242
243                 err = dccp_transmit_skb(sk, skb);
244                 ccid_hc_tx_packet_sent(dp->dccps_hc_tx_ccid, sk, 0, len);
245         } else
246                 kfree_skb(skb);
247
248         return err;
249 }
250
251 int dccp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
252 {
253         if (inet_sk_rebuild_header(sk) != 0)
254                 return -EHOSTUNREACH; /* Routing failure or similar. */
255
256         return dccp_transmit_skb(sk, (skb_cloned(skb) ?
257                                       pskb_copy(skb, GFP_ATOMIC):
258                                       skb_clone(skb, GFP_ATOMIC)));
259 }
260
261 struct sk_buff *dccp_make_response(struct sock *sk, struct dst_entry *dst,
262                                    struct request_sock *req)
263 {
264         struct dccp_hdr *dh;
265         struct dccp_request_sock *dreq;
266         const int dccp_header_size = sizeof(struct dccp_hdr) +
267                                      sizeof(struct dccp_hdr_ext) +
268                                      sizeof(struct dccp_hdr_response);
269         struct sk_buff *skb = sock_wmalloc(sk, MAX_HEADER + DCCP_MAX_OPT_LEN +
270                                                dccp_header_size, 1,
271                                            GFP_ATOMIC);
272         if (skb == NULL)
273                 return NULL;
274
275         /* Reserve space for headers. */
276         skb_reserve(skb, MAX_HEADER + DCCP_MAX_OPT_LEN + dccp_header_size);
277
278         skb->dst = dst_clone(dst);
279         skb->csum = 0;
280
281         dreq = dccp_rsk(req);
282         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_RESPONSE;
283         DCCP_SKB_CB(skb)->dccpd_seq  = dreq->dreq_iss;
284         dccp_insert_options(sk, skb);
285
286         skb->h.raw = skb_push(skb, dccp_header_size);
287
288         dh = dccp_hdr(skb);
289         memset(dh, 0, dccp_header_size);
290
291         dh->dccph_sport = inet_sk(sk)->sport;
292         dh->dccph_dport = inet_rsk(req)->rmt_port;
293         dh->dccph_doff  = (dccp_header_size +
294                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
295         dh->dccph_type  = DCCP_PKT_RESPONSE;
296         dh->dccph_x     = 1;
297         dccp_hdr_set_seq(dh, dreq->dreq_iss);
298         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dreq->dreq_isr);
299         dccp_hdr_response(skb)->dccph_resp_service = dreq->dreq_service;
300
301         dh->dccph_checksum = dccp_v4_checksum(skb, inet_rsk(req)->loc_addr,
302                                               inet_rsk(req)->rmt_addr);
303
304         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
305         return skb;
306 }
307
308 struct sk_buff *dccp_make_reset(struct sock *sk, struct dst_entry *dst,
309                                 const enum dccp_reset_codes code)
310                                    
311 {
312         struct dccp_hdr *dh;
313         struct dccp_sock *dp = dccp_sk(sk);
314         const int dccp_header_size = sizeof(struct dccp_hdr) +
315                                      sizeof(struct dccp_hdr_ext) +
316                                      sizeof(struct dccp_hdr_reset);
317         struct sk_buff *skb = sock_wmalloc(sk, MAX_HEADER + DCCP_MAX_OPT_LEN +
318                                                dccp_header_size, 1,
319                                            GFP_ATOMIC);
320         if (skb == NULL)
321                 return NULL;
322
323         /* Reserve space for headers. */
324         skb_reserve(skb, MAX_HEADER + DCCP_MAX_OPT_LEN + dccp_header_size);
325
326         skb->dst = dst_clone(dst);
327         skb->csum = 0;
328
329         dccp_inc_seqno(&dp->dccps_gss);
330
331         DCCP_SKB_CB(skb)->dccpd_reset_code = code;
332         DCCP_SKB_CB(skb)->dccpd_type       = DCCP_PKT_RESET;
333         DCCP_SKB_CB(skb)->dccpd_seq        = dp->dccps_gss;
334         dccp_insert_options(sk, skb);
335
336         skb->h.raw = skb_push(skb, dccp_header_size);
337
338         dh = dccp_hdr(skb);
339         memset(dh, 0, dccp_header_size);
340
341         dh->dccph_sport = inet_sk(sk)->sport;
342         dh->dccph_dport = inet_sk(sk)->dport;
343         dh->dccph_doff  = (dccp_header_size +
344                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
345         dh->dccph_type  = DCCP_PKT_RESET;
346         dh->dccph_x     = 1;
347         dccp_hdr_set_seq(dh, dp->dccps_gss);
348         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dp->dccps_gsr);
349
350         dccp_hdr_reset(skb)->dccph_reset_code = code;
351
352         dh->dccph_checksum = dccp_v4_checksum(skb, inet_sk(sk)->saddr,
353                                               inet_sk(sk)->daddr);
354
355         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
356         return skb;
357 }
358
359 /*
360  * Do all connect socket setups that can be done AF independent.
361  */
362 static inline void dccp_connect_init(struct sock *sk)
363 {
364         struct dst_entry *dst = __sk_dst_get(sk);
365         struct inet_connection_sock *icsk = inet_csk(sk);
366
367         sk->sk_err = 0;
368         sock_reset_flag(sk, SOCK_DONE);
369         
370         dccp_sync_mss(sk, dst_mtu(dst));
371
372         /*
373          * FIXME: set dp->{dccps_swh,dccps_swl}, with
374          * something like dccp_inc_seq
375          */
376
377         icsk->icsk_retransmits = 0;
378 }
379
380 int dccp_connect(struct sock *sk)
381 {
382         struct sk_buff *skb;
383         struct inet_connection_sock *icsk = inet_csk(sk);
384
385         dccp_connect_init(sk);
386
387         skb = alloc_skb(MAX_DCCP_HEADER + 15, sk->sk_allocation);
388         if (unlikely(skb == NULL))
389                 return -ENOBUFS;
390
391         /* Reserve space for headers. */
392         skb_reserve(skb, MAX_DCCP_HEADER);
393
394         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_REQUEST;
395         skb->csum = 0;
396         skb_set_owner_w(skb, sk);
397
398         BUG_TRAP(sk->sk_send_head == NULL);
399         sk->sk_send_head = skb;
400         dccp_transmit_skb(sk, skb_clone(skb, GFP_KERNEL));
401         DCCP_INC_STATS(DCCP_MIB_ACTIVEOPENS);
402
403         /* Timer for repeating the REQUEST until an answer. */
404         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
405                                   icsk->icsk_rto, DCCP_RTO_MAX);
406         return 0;
407 }
408
409 void dccp_send_ack(struct sock *sk)
410 {
411         /* If we have been reset, we may not send again. */
412         if (sk->sk_state != DCCP_CLOSED) {
413                 struct sk_buff *skb = alloc_skb(MAX_DCCP_HEADER, GFP_ATOMIC);
414
415                 if (skb == NULL) {
416                         inet_csk_schedule_ack(sk);
417                         inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
418                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
419                                                   TCP_DELACK_MAX,
420                                                   DCCP_RTO_MAX);
421                         return;
422                 }
423
424                 /* Reserve space for headers */
425                 skb_reserve(skb, MAX_DCCP_HEADER);
426                 skb->csum = 0;
427                 DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_ACK;
428                 skb_set_owner_w(skb, sk);
429                 dccp_transmit_skb(sk, skb);
430         }
431 }
432
433 EXPORT_SYMBOL_GPL(dccp_send_ack);
434
435 void dccp_send_delayed_ack(struct sock *sk)
436 {
437         struct inet_connection_sock *icsk = inet_csk(sk);
438         /*
439          * FIXME: tune this timer. elapsed time fixes the skew, so no problem
440          * with using 2s, and active senders also piggyback the ACK into a
441          * DATAACK packet, so this is really for quiescent senders.
442          */
443         unsigned long timeout = jiffies + 2 * HZ;
444
445         /* Use new timeout only if there wasn't a older one earlier. */
446         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
447                 /* If delack timer was blocked or is about to expire,
448                  * send ACK now.
449                  *
450                  * FIXME: check the "about to expire" part
451                  */
452                 if (icsk->icsk_ack.blocked) {
453                         dccp_send_ack(sk);
454                         return;
455                 }
456
457                 if (!time_before(timeout, icsk->icsk_ack.timeout))
458                         timeout = icsk->icsk_ack.timeout;
459         }
460         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
461         icsk->icsk_ack.timeout = timeout;
462         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
463 }
464
465 void dccp_send_sync(struct sock *sk, const u64 seq,
466                     const enum dccp_pkt_type pkt_type)
467 {
468         /*
469          * We are not putting this on the write queue, so
470          * dccp_transmit_skb() will set the ownership to this
471          * sock.
472          */
473         struct sk_buff *skb = alloc_skb(MAX_DCCP_HEADER, GFP_ATOMIC);
474
475         if (skb == NULL)
476                 /* FIXME: how to make sure the sync is sent? */
477                 return;
478
479         /* Reserve space for headers and prepare control bits. */
480         skb_reserve(skb, MAX_DCCP_HEADER);
481         skb->csum = 0;
482         DCCP_SKB_CB(skb)->dccpd_type = pkt_type;
483         DCCP_SKB_CB(skb)->dccpd_seq = seq;
484
485         skb_set_owner_w(skb, sk);
486         dccp_transmit_skb(sk, skb);
487 }
488
489 /*
490  * Send a DCCP_PKT_CLOSE/CLOSEREQ. The caller locks the socket for us. This
491  * cannot be allowed to fail queueing a DCCP_PKT_CLOSE/CLOSEREQ frame under
492  * any circumstances.
493  */
494 void dccp_send_close(struct sock *sk, const int active)
495 {
496         struct dccp_sock *dp = dccp_sk(sk);
497         struct sk_buff *skb;
498         const gfp_t prio = active ? GFP_KERNEL : GFP_ATOMIC;
499
500         skb = alloc_skb(sk->sk_prot->max_header, prio);
501         if (skb == NULL)
502                 return;
503
504         /* Reserve space for headers and prepare control bits. */
505         skb_reserve(skb, sk->sk_prot->max_header);
506         skb->csum = 0;
507         DCCP_SKB_CB(skb)->dccpd_type = dp->dccps_role == DCCP_ROLE_CLIENT ?
508                                         DCCP_PKT_CLOSE : DCCP_PKT_CLOSEREQ;
509
510         skb_set_owner_w(skb, sk);
511         if (active) {
512                 BUG_TRAP(sk->sk_send_head == NULL);
513                 sk->sk_send_head = skb;
514                 dccp_transmit_skb(sk, skb_clone(skb, prio));
515         } else
516                 dccp_transmit_skb(sk, skb);
517 }