Merge master.kernel.org:/pub/scm/linux/kernel/git/davem/net-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 "ccid.h"
20 #include "dccp.h"
21
22 static inline void dccp_event_ack_sent(struct sock *sk)
23 {
24         inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
25 }
26
27 /*
28  * All SKB's seen here are completely headerless. It is our
29  * job to build the DCCP header, and pass the packet down to
30  * IP so it can do the same plus pass the packet off to the
31  * device.
32  */
33 int dccp_transmit_skb(struct sock *sk, struct sk_buff *skb)
34 {
35         if (likely(skb != NULL)) {
36                 const struct inet_sock *inet = inet_sk(sk);
37                 struct dccp_sock *dp = dccp_sk(sk);
38                 struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
39                 struct dccp_hdr *dh;
40                 /* XXX For now we're using only 48 bits sequence numbers */
41                 const int dccp_header_size = sizeof(*dh) +
42                                              sizeof(struct dccp_hdr_ext) +
43                                           dccp_packet_hdr_len(dcb->dccpd_type);
44                 int err, set_ack = 1;
45                 u64 ackno = dp->dccps_gsr;
46
47                 dccp_inc_seqno(&dp->dccps_gss);
48
49                 switch (dcb->dccpd_type) {
50                 case DCCP_PKT_DATA:
51                         set_ack = 0;
52                         break;
53                 case DCCP_PKT_SYNC:
54                 case DCCP_PKT_SYNCACK:
55                         ackno = dcb->dccpd_seq;
56                         break;
57                 }
58
59                 dcb->dccpd_seq = dp->dccps_gss;
60                 dccp_insert_options(sk, skb);
61                 
62                 skb->h.raw = skb_push(skb, dccp_header_size);
63                 dh = dccp_hdr(skb);
64                 /*
65                  * Data packets are not cloned as they are never retransmitted
66                  */
67                 if (skb_cloned(skb))
68                         skb_set_owner_w(skb, sk);
69
70                 /* Build DCCP header and checksum it. */
71                 memset(dh, 0, dccp_header_size);
72                 dh->dccph_type  = dcb->dccpd_type;
73                 dh->dccph_sport = inet->sport;
74                 dh->dccph_dport = inet->dport;
75                 dh->dccph_doff  = (dccp_header_size + dcb->dccpd_opt_len) / 4;
76                 dh->dccph_ccval = dcb->dccpd_ccval;
77                 /* XXX For now we're using only 48 bits sequence numbers */
78                 dh->dccph_x     = 1;
79
80                 dp->dccps_awh = dp->dccps_gss;
81                 dccp_hdr_set_seq(dh, dp->dccps_gss);
82                 if (set_ack)
83                         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), ackno);
84
85                 switch (dcb->dccpd_type) {
86                 case DCCP_PKT_REQUEST:
87                         dccp_hdr_request(skb)->dccph_req_service =
88                                                         dp->dccps_service;
89                         break;
90                 case DCCP_PKT_RESET:
91                         dccp_hdr_reset(skb)->dccph_reset_code =
92                                                         dcb->dccpd_reset_code;
93                         break;
94                 }
95
96                 dh->dccph_checksum = dccp_v4_checksum(skb, inet->saddr,
97                                                       inet->daddr);
98
99                 if (set_ack)
100                         dccp_event_ack_sent(sk);
101
102                 DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
103
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                 const struct dccp_ackpkts *ap = dp->dccps_hc_rx_ackpkts;
229                 struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
230                 const int len = skb->len;
231
232                 if (sk->sk_state == DCCP_PARTOPEN) {
233                         /* See 8.1.5.  Handshake Completion */
234                         inet_csk_schedule_ack(sk);
235                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
236                                                   inet_csk(sk)->icsk_rto,
237                                                   DCCP_RTO_MAX);
238                         dcb->dccpd_type = DCCP_PKT_DATAACK;
239                         /*
240                          * FIXME: we really should have a
241                          * dccps_ack_pending or use icsk.
242                          */
243                 } else if (inet_csk_ack_scheduled(sk) ||
244                            dp->dccps_timestamp_echo != 0 ||
245                            (dp->dccps_options.dccpo_send_ack_vector &&
246                             ap->dccpap_buf_ackno != DCCP_MAX_SEQNO + 1 &&
247                             ap->dccpap_ack_seqno == DCCP_MAX_SEQNO + 1))
248                         dcb->dccpd_type = DCCP_PKT_DATAACK;
249                 else
250                         dcb->dccpd_type = DCCP_PKT_DATA;
251
252                 err = dccp_transmit_skb(sk, skb);
253                 ccid_hc_tx_packet_sent(dp->dccps_hc_tx_ccid, sk, 0, len);
254         }
255
256         return err;
257 }
258
259 int dccp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
260 {
261         if (inet_sk_rebuild_header(sk) != 0)
262                 return -EHOSTUNREACH; /* Routing failure or similar. */
263
264         return dccp_transmit_skb(sk, (skb_cloned(skb) ?
265                                       pskb_copy(skb, GFP_ATOMIC):
266                                       skb_clone(skb, GFP_ATOMIC)));
267 }
268
269 struct sk_buff *dccp_make_response(struct sock *sk, struct dst_entry *dst,
270                                    struct request_sock *req)
271 {
272         struct dccp_hdr *dh;
273         struct dccp_request_sock *dreq;
274         const int dccp_header_size = sizeof(struct dccp_hdr) +
275                                      sizeof(struct dccp_hdr_ext) +
276                                      sizeof(struct dccp_hdr_response);
277         struct sk_buff *skb = sock_wmalloc(sk, MAX_HEADER + DCCP_MAX_OPT_LEN +
278                                                dccp_header_size, 1,
279                                            GFP_ATOMIC);
280         if (skb == NULL)
281                 return NULL;
282
283         /* Reserve space for headers. */
284         skb_reserve(skb, MAX_HEADER + DCCP_MAX_OPT_LEN + dccp_header_size);
285
286         skb->dst = dst_clone(dst);
287         skb->csum = 0;
288
289         dreq = dccp_rsk(req);
290         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_RESPONSE;
291         DCCP_SKB_CB(skb)->dccpd_seq  = dreq->dreq_iss;
292         dccp_insert_options(sk, skb);
293
294         skb->h.raw = skb_push(skb, dccp_header_size);
295
296         dh = dccp_hdr(skb);
297         memset(dh, 0, dccp_header_size);
298
299         dh->dccph_sport = inet_sk(sk)->sport;
300         dh->dccph_dport = inet_rsk(req)->rmt_port;
301         dh->dccph_doff  = (dccp_header_size +
302                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
303         dh->dccph_type  = DCCP_PKT_RESPONSE;
304         dh->dccph_x     = 1;
305         dccp_hdr_set_seq(dh, dreq->dreq_iss);
306         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dreq->dreq_isr);
307         dccp_hdr_response(skb)->dccph_resp_service = dreq->dreq_service;
308
309         dh->dccph_checksum = dccp_v4_checksum(skb, inet_rsk(req)->loc_addr,
310                                               inet_rsk(req)->rmt_addr);
311
312         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
313         return skb;
314 }
315
316 struct sk_buff *dccp_make_reset(struct sock *sk, struct dst_entry *dst,
317                                 const enum dccp_reset_codes code)
318                                    
319 {
320         struct dccp_hdr *dh;
321         struct dccp_sock *dp = dccp_sk(sk);
322         const int dccp_header_size = sizeof(struct dccp_hdr) +
323                                      sizeof(struct dccp_hdr_ext) +
324                                      sizeof(struct dccp_hdr_reset);
325         struct sk_buff *skb = sock_wmalloc(sk, MAX_HEADER + DCCP_MAX_OPT_LEN +
326                                                dccp_header_size, 1,
327                                            GFP_ATOMIC);
328         if (skb == NULL)
329                 return NULL;
330
331         /* Reserve space for headers. */
332         skb_reserve(skb, MAX_HEADER + DCCP_MAX_OPT_LEN + dccp_header_size);
333
334         skb->dst = dst_clone(dst);
335         skb->csum = 0;
336
337         dccp_inc_seqno(&dp->dccps_gss);
338
339         DCCP_SKB_CB(skb)->dccpd_reset_code = code;
340         DCCP_SKB_CB(skb)->dccpd_type       = DCCP_PKT_RESET;
341         DCCP_SKB_CB(skb)->dccpd_seq        = dp->dccps_gss;
342         dccp_insert_options(sk, skb);
343
344         skb->h.raw = skb_push(skb, dccp_header_size);
345
346         dh = dccp_hdr(skb);
347         memset(dh, 0, dccp_header_size);
348
349         dh->dccph_sport = inet_sk(sk)->sport;
350         dh->dccph_dport = inet_sk(sk)->dport;
351         dh->dccph_doff  = (dccp_header_size +
352                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
353         dh->dccph_type  = DCCP_PKT_RESET;
354         dh->dccph_x     = 1;
355         dccp_hdr_set_seq(dh, dp->dccps_gss);
356         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dp->dccps_gsr);
357
358         dccp_hdr_reset(skb)->dccph_reset_code = code;
359
360         dh->dccph_checksum = dccp_v4_checksum(skb, inet_sk(sk)->saddr,
361                                               inet_sk(sk)->daddr);
362
363         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
364         return skb;
365 }
366
367 /*
368  * Do all connect socket setups that can be done AF independent.
369  */
370 static inline void dccp_connect_init(struct sock *sk)
371 {
372         struct dst_entry *dst = __sk_dst_get(sk);
373         struct inet_connection_sock *icsk = inet_csk(sk);
374
375         sk->sk_err = 0;
376         sock_reset_flag(sk, SOCK_DONE);
377         
378         dccp_sync_mss(sk, dst_mtu(dst));
379
380         /*
381          * FIXME: set dp->{dccps_swh,dccps_swl}, with
382          * something like dccp_inc_seq
383          */
384
385         icsk->icsk_retransmits = 0;
386 }
387
388 int dccp_connect(struct sock *sk)
389 {
390         struct sk_buff *skb;
391         struct inet_connection_sock *icsk = inet_csk(sk);
392
393         dccp_connect_init(sk);
394
395         skb = alloc_skb(MAX_DCCP_HEADER + 15, sk->sk_allocation);
396         if (unlikely(skb == NULL))
397                 return -ENOBUFS;
398
399         /* Reserve space for headers. */
400         skb_reserve(skb, MAX_DCCP_HEADER);
401
402         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_REQUEST;
403         skb->csum = 0;
404         skb_set_owner_w(skb, sk);
405
406         BUG_TRAP(sk->sk_send_head == NULL);
407         sk->sk_send_head = skb;
408         dccp_transmit_skb(sk, skb_clone(skb, GFP_KERNEL));
409         DCCP_INC_STATS(DCCP_MIB_ACTIVEOPENS);
410
411         /* Timer for repeating the REQUEST until an answer. */
412         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
413                                   icsk->icsk_rto, DCCP_RTO_MAX);
414         return 0;
415 }
416
417 void dccp_send_ack(struct sock *sk)
418 {
419         /* If we have been reset, we may not send again. */
420         if (sk->sk_state != DCCP_CLOSED) {
421                 struct sk_buff *skb = alloc_skb(MAX_DCCP_HEADER, GFP_ATOMIC);
422
423                 if (skb == NULL) {
424                         inet_csk_schedule_ack(sk);
425                         inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
426                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
427                                                   TCP_DELACK_MAX,
428                                                   DCCP_RTO_MAX);
429                         return;
430                 }
431
432                 /* Reserve space for headers */
433                 skb_reserve(skb, MAX_DCCP_HEADER);
434                 skb->csum = 0;
435                 DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_ACK;
436                 skb_set_owner_w(skb, sk);
437                 dccp_transmit_skb(sk, skb);
438         }
439 }
440
441 EXPORT_SYMBOL_GPL(dccp_send_ack);
442
443 void dccp_send_delayed_ack(struct sock *sk)
444 {
445         struct inet_connection_sock *icsk = inet_csk(sk);
446         /*
447          * FIXME: tune this timer. elapsed time fixes the skew, so no problem
448          * with using 2s, and active senders also piggyback the ACK into a
449          * DATAACK packet, so this is really for quiescent senders.
450          */
451         unsigned long timeout = jiffies + 2 * HZ;
452
453         /* Use new timeout only if there wasn't a older one earlier. */
454         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
455                 /* If delack timer was blocked or is about to expire,
456                  * send ACK now.
457                  *
458                  * FIXME: check the "about to expire" part
459                  */
460                 if (icsk->icsk_ack.blocked) {
461                         dccp_send_ack(sk);
462                         return;
463                 }
464
465                 if (!time_before(timeout, icsk->icsk_ack.timeout))
466                         timeout = icsk->icsk_ack.timeout;
467         }
468         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
469         icsk->icsk_ack.timeout = timeout;
470         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
471 }
472
473 void dccp_send_sync(struct sock *sk, const u64 seq,
474                     const enum dccp_pkt_type pkt_type)
475 {
476         /*
477          * We are not putting this on the write queue, so
478          * dccp_transmit_skb() will set the ownership to this
479          * sock.
480          */
481         struct sk_buff *skb = alloc_skb(MAX_DCCP_HEADER, GFP_ATOMIC);
482
483         if (skb == NULL)
484                 /* FIXME: how to make sure the sync is sent? */
485                 return;
486
487         /* Reserve space for headers and prepare control bits. */
488         skb_reserve(skb, MAX_DCCP_HEADER);
489         skb->csum = 0;
490         DCCP_SKB_CB(skb)->dccpd_type = pkt_type;
491         DCCP_SKB_CB(skb)->dccpd_seq = seq;
492
493         skb_set_owner_w(skb, sk);
494         dccp_transmit_skb(sk, skb);
495 }
496
497 /*
498  * Send a DCCP_PKT_CLOSE/CLOSEREQ. The caller locks the socket for us. This
499  * cannot be allowed to fail queueing a DCCP_PKT_CLOSE/CLOSEREQ frame under
500  * any circumstances.
501  */
502 void dccp_send_close(struct sock *sk, const int active)
503 {
504         struct dccp_sock *dp = dccp_sk(sk);
505         struct sk_buff *skb;
506         const unsigned int prio = active ? GFP_KERNEL : GFP_ATOMIC;
507
508         skb = alloc_skb(sk->sk_prot->max_header, prio);
509         if (skb == NULL)
510                 return;
511
512         /* Reserve space for headers and prepare control bits. */
513         skb_reserve(skb, sk->sk_prot->max_header);
514         skb->csum = 0;
515         DCCP_SKB_CB(skb)->dccpd_type = dp->dccps_role == DCCP_ROLE_CLIENT ?
516                                         DCCP_PKT_CLOSE : DCCP_PKT_CLOSEREQ;
517
518         skb_set_owner_w(skb, sk);
519         if (active) {
520                 BUG_TRAP(sk->sk_send_head == NULL);
521                 sk->sk_send_head = skb;
522                 dccp_transmit_skb(sk, skb_clone(skb, prio));
523         } else
524                 dccp_transmit_skb(sk, skb);
525 }