[TCP/DCCP]: Introduce net_xmit_eval
[safe/jmp/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/dccp.h>
14 #include <linux/kernel.h>
15 #include <linux/skbuff.h>
16
17 #include <net/inet_sock.h>
18 #include <net/sock.h>
19
20 #include "ackvec.h"
21 #include "ccid.h"
22 #include "dccp.h"
23
24 static inline void dccp_event_ack_sent(struct sock *sk)
25 {
26         inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
27 }
28
29 static void dccp_skb_entail(struct sock *sk, struct sk_buff *skb)
30 {
31         skb_set_owner_w(skb, sk);
32         WARN_ON(sk->sk_send_head);
33         sk->sk_send_head = skb;
34 }
35
36 /*
37  * All SKB's seen here are completely headerless. It is our
38  * job to build the DCCP header, and pass the packet down to
39  * IP so it can do the same plus pass the packet off to the
40  * device.
41  */
42 static int dccp_transmit_skb(struct sock *sk, struct sk_buff *skb)
43 {
44         if (likely(skb != NULL)) {
45                 const struct inet_sock *inet = inet_sk(sk);
46                 const struct inet_connection_sock *icsk = inet_csk(sk);
47                 struct dccp_sock *dp = dccp_sk(sk);
48                 struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
49                 struct dccp_hdr *dh;
50                 /* XXX For now we're using only 48 bits sequence numbers */
51                 const u32 dccp_header_size = sizeof(*dh) +
52                                              sizeof(struct dccp_hdr_ext) +
53                                           dccp_packet_hdr_len(dcb->dccpd_type);
54                 int err, set_ack = 1;
55                 u64 ackno = dp->dccps_gsr;
56
57                 dccp_inc_seqno(&dp->dccps_gss);
58
59                 switch (dcb->dccpd_type) {
60                 case DCCP_PKT_DATA:
61                         set_ack = 0;
62                         /* fall through */
63                 case DCCP_PKT_DATAACK:
64                         break;
65
66                 case DCCP_PKT_REQUEST:
67                         set_ack = 0;
68                         /* fall through */
69
70                 case DCCP_PKT_SYNC:
71                 case DCCP_PKT_SYNCACK:
72                         ackno = dcb->dccpd_seq;
73                         /* fall through */
74                 default:
75                         /*
76                          * Only data packets should come through with skb->sk
77                          * set.
78                          */
79                         WARN_ON(skb->sk);
80                         skb_set_owner_w(skb, sk);
81                         break;
82                 }
83
84                 dcb->dccpd_seq = dp->dccps_gss;
85
86                 if (dccp_insert_options(sk, skb)) {
87                         kfree_skb(skb);
88                         return -EPROTO;
89                 }
90                 
91
92                 /* Build DCCP header and checksum it. */
93                 dh = dccp_zeroed_hdr(skb, dccp_header_size);
94                 dh->dccph_type  = dcb->dccpd_type;
95                 dh->dccph_sport = inet->sport;
96                 dh->dccph_dport = inet->dport;
97                 dh->dccph_doff  = (dccp_header_size + dcb->dccpd_opt_len) / 4;
98                 dh->dccph_ccval = dcb->dccpd_ccval;
99                 dh->dccph_cscov = dp->dccps_pcslen;
100                 /* XXX For now we're using only 48 bits sequence numbers */
101                 dh->dccph_x     = 1;
102
103                 dp->dccps_awh = dp->dccps_gss;
104                 dccp_hdr_set_seq(dh, dp->dccps_gss);
105                 if (set_ack)
106                         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), ackno);
107
108                 switch (dcb->dccpd_type) {
109                 case DCCP_PKT_REQUEST:
110                         dccp_hdr_request(skb)->dccph_req_service =
111                                                         dp->dccps_service;
112                         break;
113                 case DCCP_PKT_RESET:
114                         dccp_hdr_reset(skb)->dccph_reset_code =
115                                                         dcb->dccpd_reset_code;
116                         break;
117                 }
118
119                 icsk->icsk_af_ops->send_check(sk, 0, skb);
120
121                 if (set_ack)
122                         dccp_event_ack_sent(sk);
123
124                 DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
125
126                 memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
127                 err = icsk->icsk_af_ops->queue_xmit(skb, sk, 0);
128                 return net_xmit_eval(err);
129         }
130         return -ENOBUFS;
131 }
132
133 unsigned int dccp_sync_mss(struct sock *sk, u32 pmtu)
134 {
135         struct inet_connection_sock *icsk = inet_csk(sk);
136         struct dccp_sock *dp = dccp_sk(sk);
137         int mss_now = (pmtu - icsk->icsk_af_ops->net_header_len -
138                        sizeof(struct dccp_hdr) - sizeof(struct dccp_hdr_ext));
139
140         /* Now subtract optional transport overhead */
141         mss_now -= icsk->icsk_ext_hdr_len;
142
143         /*
144          * FIXME: this should come from the CCID infrastructure, where, say,
145          * TFRC will say it wants TIMESTAMPS, ELAPSED time, etc, for now lets
146          * put a rough estimate for NDP + TIMESTAMP + TIMESTAMP_ECHO + ELAPSED
147          * TIME + TFRC_OPT_LOSS_EVENT_RATE + TFRC_OPT_RECEIVE_RATE + padding to
148          * make it a multiple of 4
149          */
150
151         mss_now -= ((5 + 6 + 10 + 6 + 6 + 6 + 3) / 4) * 4;
152
153         /* And store cached results */
154         icsk->icsk_pmtu_cookie = pmtu;
155         dp->dccps_mss_cache = mss_now;
156
157         return mss_now;
158 }
159
160 EXPORT_SYMBOL_GPL(dccp_sync_mss);
161
162 void dccp_write_space(struct sock *sk)
163 {
164         read_lock(&sk->sk_callback_lock);
165
166         if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
167                 wake_up_interruptible(sk->sk_sleep);
168         /* Should agree with poll, otherwise some programs break */
169         if (sock_writeable(sk))
170                 sk_wake_async(sk, 2, POLL_OUT);
171
172         read_unlock(&sk->sk_callback_lock);
173 }
174
175 /**
176  * dccp_wait_for_ccid - Wait for ccid to tell us we can send a packet
177  * @sk: socket to wait for
178  * @timeo: for how long
179  */
180 static int dccp_wait_for_ccid(struct sock *sk, struct sk_buff *skb,
181                               long *timeo)
182 {
183         struct dccp_sock *dp = dccp_sk(sk);
184         DEFINE_WAIT(wait);
185         long delay;
186         int rc;
187
188         while (1) {
189                 prepare_to_wait(sk->sk_sleep, &wait, TASK_INTERRUPTIBLE);
190
191                 if (sk->sk_err)
192                         goto do_error;
193                 if (!*timeo)
194                         goto do_nonblock;
195                 if (signal_pending(current))
196                         goto do_interrupted;
197
198                 rc = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
199                                             skb->len);
200                 if (rc <= 0)
201                         break;
202                 delay = msecs_to_jiffies(rc);
203                 if (delay > *timeo || delay < 0)
204                         goto do_nonblock;
205
206                 sk->sk_write_pending++;
207                 release_sock(sk);
208                 *timeo -= schedule_timeout(delay);
209                 lock_sock(sk);
210                 sk->sk_write_pending--;
211         }
212 out:
213         finish_wait(sk->sk_sleep, &wait);
214         return rc;
215
216 do_error:
217         rc = -EPIPE;
218         goto out;
219 do_nonblock:
220         rc = -EAGAIN;
221         goto out;
222 do_interrupted:
223         rc = sock_intr_errno(*timeo);
224         goto out;
225 }
226
227 static void dccp_write_xmit_timer(unsigned long data) {
228         struct sock *sk = (struct sock *)data;
229         struct dccp_sock *dp = dccp_sk(sk);
230
231         bh_lock_sock(sk);
232         if (sock_owned_by_user(sk))
233                 sk_reset_timer(sk, &dp->dccps_xmit_timer, jiffies+1);
234         else
235                 dccp_write_xmit(sk, 0);
236         bh_unlock_sock(sk);
237         sock_put(sk);
238 }
239
240 void dccp_write_xmit(struct sock *sk, int block)
241 {
242         struct dccp_sock *dp = dccp_sk(sk);
243         struct sk_buff *skb;
244         long timeo = DCCP_XMIT_TIMEO;   /* If a packet is taking longer than
245                                            this we have other issues */
246
247         while ((skb = skb_peek(&sk->sk_write_queue))) {
248                 int err = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
249                                          skb->len);
250
251                 if (err > 0) {
252                         if (!block) {
253                                 sk_reset_timer(sk, &dp->dccps_xmit_timer,
254                                                 msecs_to_jiffies(err)+jiffies);
255                                 break;
256                         } else {
257                                 err = dccp_wait_for_ccid(sk, skb, &timeo);
258                                 timeo = DCCP_XMIT_TIMEO;
259                         }
260                         if (err) {
261                                 printk(KERN_CRIT "%s:err at dccp_wait_for_ccid"
262                                                  " %d\n", __FUNCTION__, err);
263                                 dump_stack();
264                         }
265                 }
266
267                 skb_dequeue(&sk->sk_write_queue);
268                 if (err == 0) {
269                         struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
270                         const int len = skb->len;
271
272                         if (sk->sk_state == DCCP_PARTOPEN) {
273                                 /* See 8.1.5.  Handshake Completion */
274                                 inet_csk_schedule_ack(sk);
275                                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
276                                                   inet_csk(sk)->icsk_rto,
277                                                   DCCP_RTO_MAX);
278                                 dcb->dccpd_type = DCCP_PKT_DATAACK;
279                         } else if (dccp_ack_pending(sk))
280                                 dcb->dccpd_type = DCCP_PKT_DATAACK;
281                         else
282                                 dcb->dccpd_type = DCCP_PKT_DATA;
283
284                         err = dccp_transmit_skb(sk, skb);
285                         ccid_hc_tx_packet_sent(dp->dccps_hc_tx_ccid, sk, 0, len);
286                         if (err) {
287                                 printk(KERN_CRIT "%s:err from "
288                                                  "ccid_hc_tx_packet_sent %d\n",
289                                                  __FUNCTION__, err);
290                                 dump_stack();
291                         }
292                 } else
293                         kfree(skb);
294         }
295 }
296
297 int dccp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
298 {
299         if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk) != 0)
300                 return -EHOSTUNREACH; /* Routing failure or similar. */
301
302         return dccp_transmit_skb(sk, (skb_cloned(skb) ?
303                                       pskb_copy(skb, GFP_ATOMIC):
304                                       skb_clone(skb, GFP_ATOMIC)));
305 }
306
307 struct sk_buff *dccp_make_response(struct sock *sk, struct dst_entry *dst,
308                                    struct request_sock *req)
309 {
310         struct dccp_hdr *dh;
311         struct dccp_request_sock *dreq;
312         const u32 dccp_header_size = sizeof(struct dccp_hdr) +
313                                      sizeof(struct dccp_hdr_ext) +
314                                      sizeof(struct dccp_hdr_response);
315         struct sk_buff *skb = sock_wmalloc(sk, sk->sk_prot->max_header, 1,
316                                            GFP_ATOMIC);
317         if (skb == NULL)
318                 return NULL;
319
320         /* Reserve space for headers. */
321         skb_reserve(skb, sk->sk_prot->max_header);
322
323         skb->dst = dst_clone(dst);
324
325         dreq = dccp_rsk(req);
326         if (inet_rsk(req)->acked)       /* increase ISS upon retransmission */
327                 dccp_inc_seqno(&dreq->dreq_iss);
328         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_RESPONSE;
329         DCCP_SKB_CB(skb)->dccpd_seq  = dreq->dreq_iss;
330
331         if (dccp_insert_options(sk, skb)) {
332                 kfree_skb(skb);
333                 return NULL;
334         }
335
336         dh = dccp_zeroed_hdr(skb, dccp_header_size);
337
338         dh->dccph_sport = inet_sk(sk)->sport;
339         dh->dccph_dport = inet_rsk(req)->rmt_port;
340         dh->dccph_doff  = (dccp_header_size +
341                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
342         dh->dccph_type  = DCCP_PKT_RESPONSE;
343         dh->dccph_x     = 1;
344         dccp_hdr_set_seq(dh, dreq->dreq_iss);
345         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dreq->dreq_isr);
346         dccp_hdr_response(skb)->dccph_resp_service = dreq->dreq_service;
347
348         dccp_csum_outgoing(skb);
349
350         /* We use `acked' to remember that a Response was already sent. */
351         inet_rsk(req)->acked = 1;
352         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
353         return skb;
354 }
355
356 EXPORT_SYMBOL_GPL(dccp_make_response);
357
358 static struct sk_buff *dccp_make_reset(struct sock *sk, struct dst_entry *dst,
359                                        const enum dccp_reset_codes code)
360                                    
361 {
362         struct dccp_hdr *dh;
363         struct dccp_sock *dp = dccp_sk(sk);
364         const u32 dccp_header_size = sizeof(struct dccp_hdr) +
365                                      sizeof(struct dccp_hdr_ext) +
366                                      sizeof(struct dccp_hdr_reset);
367         struct sk_buff *skb = sock_wmalloc(sk, sk->sk_prot->max_header, 1,
368                                            GFP_ATOMIC);
369         if (skb == NULL)
370                 return NULL;
371
372         /* Reserve space for headers. */
373         skb_reserve(skb, sk->sk_prot->max_header);
374
375         skb->dst = dst_clone(dst);
376
377         dccp_inc_seqno(&dp->dccps_gss);
378
379         DCCP_SKB_CB(skb)->dccpd_reset_code = code;
380         DCCP_SKB_CB(skb)->dccpd_type       = DCCP_PKT_RESET;
381         DCCP_SKB_CB(skb)->dccpd_seq        = dp->dccps_gss;
382
383         if (dccp_insert_options(sk, skb)) {
384                 kfree_skb(skb);
385                 return NULL;
386         }
387
388         dh = dccp_zeroed_hdr(skb, dccp_header_size);
389
390         dh->dccph_sport = inet_sk(sk)->sport;
391         dh->dccph_dport = inet_sk(sk)->dport;
392         dh->dccph_doff  = (dccp_header_size +
393                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
394         dh->dccph_type  = DCCP_PKT_RESET;
395         dh->dccph_x     = 1;
396         dccp_hdr_set_seq(dh, dp->dccps_gss);
397         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dp->dccps_gsr);
398
399         dccp_hdr_reset(skb)->dccph_reset_code = code;
400         inet_csk(sk)->icsk_af_ops->send_check(sk, 0, skb);
401
402         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
403         return skb;
404 }
405
406 int dccp_send_reset(struct sock *sk, enum dccp_reset_codes code)
407 {
408         /*
409          * FIXME: what if rebuild_header fails?
410          * Should we be doing a rebuild_header here?
411          */
412         int err = inet_sk_rebuild_header(sk);
413
414         if (err == 0) {
415                 struct sk_buff *skb = dccp_make_reset(sk, sk->sk_dst_cache,
416                                                       code);
417                 if (skb != NULL) {
418                         memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
419                         err = inet_csk(sk)->icsk_af_ops->queue_xmit(skb, sk, 0);
420                         return net_xmit_eval(err);
421                 }
422         }
423
424         return err;
425 }
426
427 /*
428  * Do all connect socket setups that can be done AF independent.
429  */
430 static inline void dccp_connect_init(struct sock *sk)
431 {
432         struct dccp_sock *dp = dccp_sk(sk);
433         struct dst_entry *dst = __sk_dst_get(sk);
434         struct inet_connection_sock *icsk = inet_csk(sk);
435
436         sk->sk_err = 0;
437         sock_reset_flag(sk, SOCK_DONE);
438         
439         dccp_sync_mss(sk, dst_mtu(dst));
440
441         /*
442          * SWL and AWL are initially adjusted so that they are not less than
443          * the initial Sequence Numbers received and sent, respectively:
444          *      SWL := max(GSR + 1 - floor(W/4), ISR),
445          *      AWL := max(GSS - W' + 1, ISS).
446          * These adjustments MUST be applied only at the beginning of the
447          * connection.
448          */
449         dccp_update_gss(sk, dp->dccps_iss);
450         dccp_set_seqno(&dp->dccps_awl, max48(dp->dccps_awl, dp->dccps_iss));
451
452         /* S.GAR - greatest valid acknowledgement number received on a non-Sync;
453          *         initialized to S.ISS (sec. 8.5)                            */
454         dp->dccps_gar = dp->dccps_iss;
455
456         icsk->icsk_retransmits = 0;
457         init_timer(&dp->dccps_xmit_timer);
458         dp->dccps_xmit_timer.data = (unsigned long)sk;
459         dp->dccps_xmit_timer.function = dccp_write_xmit_timer;
460 }
461
462 int dccp_connect(struct sock *sk)
463 {
464         struct sk_buff *skb;
465         struct inet_connection_sock *icsk = inet_csk(sk);
466
467         dccp_connect_init(sk);
468
469         skb = alloc_skb(sk->sk_prot->max_header, sk->sk_allocation);
470         if (unlikely(skb == NULL))
471                 return -ENOBUFS;
472
473         /* Reserve space for headers. */
474         skb_reserve(skb, sk->sk_prot->max_header);
475
476         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_REQUEST;
477
478         dccp_skb_entail(sk, skb);
479         dccp_transmit_skb(sk, skb_clone(skb, GFP_KERNEL));
480         DCCP_INC_STATS(DCCP_MIB_ACTIVEOPENS);
481
482         /* Timer for repeating the REQUEST until an answer. */
483         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
484                                   icsk->icsk_rto, DCCP_RTO_MAX);
485         return 0;
486 }
487
488 EXPORT_SYMBOL_GPL(dccp_connect);
489
490 void dccp_send_ack(struct sock *sk)
491 {
492         /* If we have been reset, we may not send again. */
493         if (sk->sk_state != DCCP_CLOSED) {
494                 struct sk_buff *skb = alloc_skb(sk->sk_prot->max_header,
495                                                 GFP_ATOMIC);
496
497                 if (skb == NULL) {
498                         inet_csk_schedule_ack(sk);
499                         inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
500                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
501                                                   TCP_DELACK_MAX,
502                                                   DCCP_RTO_MAX);
503                         return;
504                 }
505
506                 /* Reserve space for headers */
507                 skb_reserve(skb, sk->sk_prot->max_header);
508                 DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_ACK;
509                 dccp_transmit_skb(sk, skb);
510         }
511 }
512
513 EXPORT_SYMBOL_GPL(dccp_send_ack);
514
515 void dccp_send_delayed_ack(struct sock *sk)
516 {
517         struct inet_connection_sock *icsk = inet_csk(sk);
518         /*
519          * FIXME: tune this timer. elapsed time fixes the skew, so no problem
520          * with using 2s, and active senders also piggyback the ACK into a
521          * DATAACK packet, so this is really for quiescent senders.
522          */
523         unsigned long timeout = jiffies + 2 * HZ;
524
525         /* Use new timeout only if there wasn't a older one earlier. */
526         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
527                 /* If delack timer was blocked or is about to expire,
528                  * send ACK now.
529                  *
530                  * FIXME: check the "about to expire" part
531                  */
532                 if (icsk->icsk_ack.blocked) {
533                         dccp_send_ack(sk);
534                         return;
535                 }
536
537                 if (!time_before(timeout, icsk->icsk_ack.timeout))
538                         timeout = icsk->icsk_ack.timeout;
539         }
540         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
541         icsk->icsk_ack.timeout = timeout;
542         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
543 }
544
545 void dccp_send_sync(struct sock *sk, const u64 seq,
546                     const enum dccp_pkt_type pkt_type)
547 {
548         /*
549          * We are not putting this on the write queue, so
550          * dccp_transmit_skb() will set the ownership to this
551          * sock.
552          */
553         struct sk_buff *skb = alloc_skb(sk->sk_prot->max_header, GFP_ATOMIC);
554
555         if (skb == NULL)
556                 /* FIXME: how to make sure the sync is sent? */
557                 return;
558
559         /* Reserve space for headers and prepare control bits. */
560         skb_reserve(skb, sk->sk_prot->max_header);
561         DCCP_SKB_CB(skb)->dccpd_type = pkt_type;
562         DCCP_SKB_CB(skb)->dccpd_seq = seq;
563
564         dccp_transmit_skb(sk, skb);
565 }
566
567 EXPORT_SYMBOL_GPL(dccp_send_sync);
568
569 /*
570  * Send a DCCP_PKT_CLOSE/CLOSEREQ. The caller locks the socket for us. This
571  * cannot be allowed to fail queueing a DCCP_PKT_CLOSE/CLOSEREQ frame under
572  * any circumstances.
573  */
574 void dccp_send_close(struct sock *sk, const int active)
575 {
576         struct dccp_sock *dp = dccp_sk(sk);
577         struct sk_buff *skb;
578         const gfp_t prio = active ? GFP_KERNEL : GFP_ATOMIC;
579
580         skb = alloc_skb(sk->sk_prot->max_header, prio);
581         if (skb == NULL)
582                 return;
583
584         /* Reserve space for headers and prepare control bits. */
585         skb_reserve(skb, sk->sk_prot->max_header);
586         DCCP_SKB_CB(skb)->dccpd_type = dp->dccps_role == DCCP_ROLE_CLIENT ?
587                                         DCCP_PKT_CLOSE : DCCP_PKT_CLOSEREQ;
588
589         if (active) {
590                 dccp_write_xmit(sk, 1);
591                 dccp_skb_entail(sk, skb);
592                 dccp_transmit_skb(sk, skb_clone(skb, prio));
593                 /* FIXME do we need a retransmit timer here? */
594         } else
595                 dccp_transmit_skb(sk, skb);
596 }