[DCCP]: Combine allocating & zeroing header space on skb
[linux-2.6.git] / 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                 /* XXX For now we're using only 48 bits sequence numbers */
100                 dh->dccph_x     = 1;
101
102                 dp->dccps_awh = dp->dccps_gss;
103                 dccp_hdr_set_seq(dh, dp->dccps_gss);
104                 if (set_ack)
105                         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), ackno);
106
107                 switch (dcb->dccpd_type) {
108                 case DCCP_PKT_REQUEST:
109                         dccp_hdr_request(skb)->dccph_req_service =
110                                                         dp->dccps_service;
111                         break;
112                 case DCCP_PKT_RESET:
113                         dccp_hdr_reset(skb)->dccph_reset_code =
114                                                         dcb->dccpd_reset_code;
115                         break;
116                 }
117
118                 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
119
120                 if (set_ack)
121                         dccp_event_ack_sent(sk);
122
123                 DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
124
125                 memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
126                 err = icsk->icsk_af_ops->queue_xmit(skb, sk, 0);
127                 if (err <= 0)
128                         return err;
129
130                 /* NET_XMIT_CN is special. It does not guarantee,
131                  * that this packet is lost. It tells that device
132                  * is about to start to drop packets or already
133                  * drops some packets of the same priority and
134                  * invokes us to send less aggressively.
135                  */
136                 return err == NET_XMIT_CN ? 0 : err;
137         }
138         return -ENOBUFS;
139 }
140
141 unsigned int dccp_sync_mss(struct sock *sk, u32 pmtu)
142 {
143         struct inet_connection_sock *icsk = inet_csk(sk);
144         struct dccp_sock *dp = dccp_sk(sk);
145         int mss_now = (pmtu - icsk->icsk_af_ops->net_header_len -
146                        sizeof(struct dccp_hdr) - sizeof(struct dccp_hdr_ext));
147
148         /* Now subtract optional transport overhead */
149         mss_now -= icsk->icsk_ext_hdr_len;
150
151         /*
152          * FIXME: this should come from the CCID infrastructure, where, say,
153          * TFRC will say it wants TIMESTAMPS, ELAPSED time, etc, for now lets
154          * put a rough estimate for NDP + TIMESTAMP + TIMESTAMP_ECHO + ELAPSED
155          * TIME + TFRC_OPT_LOSS_EVENT_RATE + TFRC_OPT_RECEIVE_RATE + padding to
156          * make it a multiple of 4
157          */
158
159         mss_now -= ((5 + 6 + 10 + 6 + 6 + 6 + 3) / 4) * 4;
160
161         /* And store cached results */
162         icsk->icsk_pmtu_cookie = pmtu;
163         dp->dccps_mss_cache = mss_now;
164
165         return mss_now;
166 }
167
168 EXPORT_SYMBOL_GPL(dccp_sync_mss);
169
170 void dccp_write_space(struct sock *sk)
171 {
172         read_lock(&sk->sk_callback_lock);
173
174         if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
175                 wake_up_interruptible(sk->sk_sleep);
176         /* Should agree with poll, otherwise some programs break */
177         if (sock_writeable(sk))
178                 sk_wake_async(sk, 2, POLL_OUT);
179
180         read_unlock(&sk->sk_callback_lock);
181 }
182
183 /**
184  * dccp_wait_for_ccid - Wait for ccid to tell us we can send a packet
185  * @sk: socket to wait for
186  * @timeo: for how long
187  */
188 static int dccp_wait_for_ccid(struct sock *sk, struct sk_buff *skb,
189                               long *timeo)
190 {
191         struct dccp_sock *dp = dccp_sk(sk);
192         DEFINE_WAIT(wait);
193         long delay;
194         int rc;
195
196         while (1) {
197                 prepare_to_wait(sk->sk_sleep, &wait, TASK_INTERRUPTIBLE);
198
199                 if (sk->sk_err)
200                         goto do_error;
201                 if (!*timeo)
202                         goto do_nonblock;
203                 if (signal_pending(current))
204                         goto do_interrupted;
205
206                 rc = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
207                                             skb->len);
208                 if (rc <= 0)
209                         break;
210                 delay = msecs_to_jiffies(rc);
211                 if (delay > *timeo || delay < 0)
212                         goto do_nonblock;
213
214                 sk->sk_write_pending++;
215                 release_sock(sk);
216                 *timeo -= schedule_timeout(delay);
217                 lock_sock(sk);
218                 sk->sk_write_pending--;
219         }
220 out:
221         finish_wait(sk->sk_sleep, &wait);
222         return rc;
223
224 do_error:
225         rc = -EPIPE;
226         goto out;
227 do_nonblock:
228         rc = -EAGAIN;
229         goto out;
230 do_interrupted:
231         rc = sock_intr_errno(*timeo);
232         goto out;
233 }
234
235 static void dccp_write_xmit_timer(unsigned long data) {
236         struct sock *sk = (struct sock *)data;
237         struct dccp_sock *dp = dccp_sk(sk);
238
239         bh_lock_sock(sk);
240         if (sock_owned_by_user(sk))
241                 sk_reset_timer(sk, &dp->dccps_xmit_timer, jiffies+1);
242         else
243                 dccp_write_xmit(sk, 0);
244         bh_unlock_sock(sk);
245         sock_put(sk);
246 }
247
248 void dccp_write_xmit(struct sock *sk, int block)
249 {
250         struct dccp_sock *dp = dccp_sk(sk);
251         struct sk_buff *skb;
252         long timeo = 30000;     /* If a packet is taking longer than 2 secs
253                                    we have other issues */
254
255         while ((skb = skb_peek(&sk->sk_write_queue))) {
256                 int err = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
257                                          skb->len);
258
259                 if (err > 0) {
260                         if (!block) {
261                                 sk_reset_timer(sk, &dp->dccps_xmit_timer,
262                                                 msecs_to_jiffies(err)+jiffies);
263                                 break;
264                         } else
265                                 err = dccp_wait_for_ccid(sk, skb, &timeo);
266                         if (err) {
267                                 printk(KERN_CRIT "%s:err at dccp_wait_for_ccid"
268                                                  " %d\n", __FUNCTION__, err);
269                                 dump_stack();
270                         }
271                 }
272
273                 skb_dequeue(&sk->sk_write_queue);
274                 if (err == 0) {
275                         struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
276                         const int len = skb->len;
277
278                         if (sk->sk_state == DCCP_PARTOPEN) {
279                                 /* See 8.1.5.  Handshake Completion */
280                                 inet_csk_schedule_ack(sk);
281                                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
282                                                   inet_csk(sk)->icsk_rto,
283                                                   DCCP_RTO_MAX);
284                                 dcb->dccpd_type = DCCP_PKT_DATAACK;
285                         } else if (dccp_ack_pending(sk))
286                                 dcb->dccpd_type = DCCP_PKT_DATAACK;
287                         else
288                                 dcb->dccpd_type = DCCP_PKT_DATA;
289
290                         err = dccp_transmit_skb(sk, skb);
291                         ccid_hc_tx_packet_sent(dp->dccps_hc_tx_ccid, sk, 0, len);
292                         if (err) {
293                                 printk(KERN_CRIT "%s:err from "
294                                                  "ccid_hc_tx_packet_sent %d\n",
295                                                  __FUNCTION__, err);
296                                 dump_stack();
297                         }
298                 } else
299                         kfree(skb);
300         }
301 }
302
303 int dccp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
304 {
305         if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk) != 0)
306                 return -EHOSTUNREACH; /* Routing failure or similar. */
307
308         return dccp_transmit_skb(sk, (skb_cloned(skb) ?
309                                       pskb_copy(skb, GFP_ATOMIC):
310                                       skb_clone(skb, GFP_ATOMIC)));
311 }
312
313 struct sk_buff *dccp_make_response(struct sock *sk, struct dst_entry *dst,
314                                    struct request_sock *req)
315 {
316         struct dccp_hdr *dh;
317         struct dccp_request_sock *dreq;
318         const u32 dccp_header_size = sizeof(struct dccp_hdr) +
319                                      sizeof(struct dccp_hdr_ext) +
320                                      sizeof(struct dccp_hdr_response);
321         struct sk_buff *skb = sock_wmalloc(sk, sk->sk_prot->max_header, 1,
322                                            GFP_ATOMIC);
323         if (skb == NULL)
324                 return NULL;
325
326         /* Reserve space for headers. */
327         skb_reserve(skb, sk->sk_prot->max_header);
328
329         skb->dst = dst_clone(dst);
330         skb->csum = 0;
331
332         dreq = dccp_rsk(req);
333         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_RESPONSE;
334         DCCP_SKB_CB(skb)->dccpd_seq  = dreq->dreq_iss;
335
336         if (dccp_insert_options(sk, skb)) {
337                 kfree_skb(skb);
338                 return NULL;
339         }
340
341         dh = dccp_zeroed_hdr(skb, dccp_header_size);
342
343         dh->dccph_sport = inet_sk(sk)->sport;
344         dh->dccph_dport = inet_rsk(req)->rmt_port;
345         dh->dccph_doff  = (dccp_header_size +
346                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
347         dh->dccph_type  = DCCP_PKT_RESPONSE;
348         dh->dccph_x     = 1;
349         dccp_hdr_set_seq(dh, dreq->dreq_iss);
350         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dreq->dreq_isr);
351         dccp_hdr_response(skb)->dccph_resp_service = dreq->dreq_service;
352
353         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
354         return skb;
355 }
356
357 EXPORT_SYMBOL_GPL(dccp_make_response);
358
359 static struct sk_buff *dccp_make_reset(struct sock *sk, struct dst_entry *dst,
360                                        const enum dccp_reset_codes code)
361                                    
362 {
363         struct dccp_hdr *dh;
364         struct dccp_sock *dp = dccp_sk(sk);
365         const u32 dccp_header_size = sizeof(struct dccp_hdr) +
366                                      sizeof(struct dccp_hdr_ext) +
367                                      sizeof(struct dccp_hdr_reset);
368         struct sk_buff *skb = sock_wmalloc(sk, sk->sk_prot->max_header, 1,
369                                            GFP_ATOMIC);
370         if (skb == NULL)
371                 return NULL;
372
373         /* Reserve space for headers. */
374         skb_reserve(skb, sk->sk_prot->max_header);
375
376         skb->dst = dst_clone(dst);
377         skb->csum = 0;
378
379         dccp_inc_seqno(&dp->dccps_gss);
380
381         DCCP_SKB_CB(skb)->dccpd_reset_code = code;
382         DCCP_SKB_CB(skb)->dccpd_type       = DCCP_PKT_RESET;
383         DCCP_SKB_CB(skb)->dccpd_seq        = dp->dccps_gss;
384
385         if (dccp_insert_options(sk, skb)) {
386                 kfree_skb(skb);
387                 return NULL;
388         }
389
390         dh = dccp_zeroed_hdr(skb, dccp_header_size);
391
392         dh->dccph_sport = inet_sk(sk)->sport;
393         dh->dccph_dport = inet_sk(sk)->dport;
394         dh->dccph_doff  = (dccp_header_size +
395                            DCCP_SKB_CB(skb)->dccpd_opt_len) / 4;
396         dh->dccph_type  = DCCP_PKT_RESET;
397         dh->dccph_x     = 1;
398         dccp_hdr_set_seq(dh, dp->dccps_gss);
399         dccp_hdr_set_ack(dccp_hdr_ack_bits(skb), dp->dccps_gsr);
400
401         dccp_hdr_reset(skb)->dccph_reset_code = code;
402         inet_csk(sk)->icsk_af_ops->send_check(sk, skb->len, skb);
403
404         DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
405         return skb;
406 }
407
408 int dccp_send_reset(struct sock *sk, enum dccp_reset_codes code)
409 {
410         /*
411          * FIXME: what if rebuild_header fails?
412          * Should we be doing a rebuild_header here?
413          */
414         int err = inet_sk_rebuild_header(sk);
415
416         if (err == 0) {
417                 struct sk_buff *skb = dccp_make_reset(sk, sk->sk_dst_cache,
418                                                       code);
419                 if (skb != NULL) {
420                         memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
421                         err = inet_csk(sk)->icsk_af_ops->queue_xmit(skb, sk, 0);
422                         if (err == NET_XMIT_CN)
423                                 err = 0;
424                 }
425         }
426
427         return err;
428 }
429
430 /*
431  * Do all connect socket setups that can be done AF independent.
432  */
433 static inline void dccp_connect_init(struct sock *sk)
434 {
435         struct dccp_sock *dp = dccp_sk(sk);
436         struct dst_entry *dst = __sk_dst_get(sk);
437         struct inet_connection_sock *icsk = inet_csk(sk);
438
439         sk->sk_err = 0;
440         sock_reset_flag(sk, SOCK_DONE);
441         
442         dccp_sync_mss(sk, dst_mtu(dst));
443
444         dccp_update_gss(sk, dp->dccps_iss);
445         /*
446          * SWL and AWL are initially adjusted so that they are not less than
447          * the initial Sequence Numbers received and sent, respectively:
448          *      SWL := max(GSR + 1 - floor(W/4), ISR),
449          *      AWL := max(GSS - W' + 1, ISS).
450          * These adjustments MUST be applied only at the beginning of the
451          * connection.
452          */
453         dccp_set_seqno(&dp->dccps_awl, max48(dp->dccps_awl, dp->dccps_iss));
454
455         icsk->icsk_retransmits = 0;
456         init_timer(&dp->dccps_xmit_timer);
457         dp->dccps_xmit_timer.data = (unsigned long)sk;
458         dp->dccps_xmit_timer.function = dccp_write_xmit_timer;
459 }
460
461 int dccp_connect(struct sock *sk)
462 {
463         struct sk_buff *skb;
464         struct inet_connection_sock *icsk = inet_csk(sk);
465
466         dccp_connect_init(sk);
467
468         skb = alloc_skb(sk->sk_prot->max_header, sk->sk_allocation);
469         if (unlikely(skb == NULL))
470                 return -ENOBUFS;
471
472         /* Reserve space for headers. */
473         skb_reserve(skb, sk->sk_prot->max_header);
474
475         DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_REQUEST;
476         skb->csum = 0;
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                 skb->csum = 0;
509                 DCCP_SKB_CB(skb)->dccpd_type = DCCP_PKT_ACK;
510                 dccp_transmit_skb(sk, skb);
511         }
512 }
513
514 EXPORT_SYMBOL_GPL(dccp_send_ack);
515
516 void dccp_send_delayed_ack(struct sock *sk)
517 {
518         struct inet_connection_sock *icsk = inet_csk(sk);
519         /*
520          * FIXME: tune this timer. elapsed time fixes the skew, so no problem
521          * with using 2s, and active senders also piggyback the ACK into a
522          * DATAACK packet, so this is really for quiescent senders.
523          */
524         unsigned long timeout = jiffies + 2 * HZ;
525
526         /* Use new timeout only if there wasn't a older one earlier. */
527         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
528                 /* If delack timer was blocked or is about to expire,
529                  * send ACK now.
530                  *
531                  * FIXME: check the "about to expire" part
532                  */
533                 if (icsk->icsk_ack.blocked) {
534                         dccp_send_ack(sk);
535                         return;
536                 }
537
538                 if (!time_before(timeout, icsk->icsk_ack.timeout))
539                         timeout = icsk->icsk_ack.timeout;
540         }
541         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
542         icsk->icsk_ack.timeout = timeout;
543         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
544 }
545
546 void dccp_send_sync(struct sock *sk, const u64 seq,
547                     const enum dccp_pkt_type pkt_type)
548 {
549         /*
550          * We are not putting this on the write queue, so
551          * dccp_transmit_skb() will set the ownership to this
552          * sock.
553          */
554         struct sk_buff *skb = alloc_skb(sk->sk_prot->max_header, GFP_ATOMIC);
555
556         if (skb == NULL)
557                 /* FIXME: how to make sure the sync is sent? */
558                 return;
559
560         /* Reserve space for headers and prepare control bits. */
561         skb_reserve(skb, sk->sk_prot->max_header);
562         skb->csum = 0;
563         DCCP_SKB_CB(skb)->dccpd_type = pkt_type;
564         DCCP_SKB_CB(skb)->dccpd_seq = seq;
565
566         dccp_transmit_skb(sk, skb);
567 }
568
569 EXPORT_SYMBOL_GPL(dccp_send_sync);
570
571 /*
572  * Send a DCCP_PKT_CLOSE/CLOSEREQ. The caller locks the socket for us. This
573  * cannot be allowed to fail queueing a DCCP_PKT_CLOSE/CLOSEREQ frame under
574  * any circumstances.
575  */
576 void dccp_send_close(struct sock *sk, const int active)
577 {
578         struct dccp_sock *dp = dccp_sk(sk);
579         struct sk_buff *skb;
580         const gfp_t prio = active ? GFP_KERNEL : GFP_ATOMIC;
581
582         skb = alloc_skb(sk->sk_prot->max_header, prio);
583         if (skb == NULL)
584                 return;
585
586         /* Reserve space for headers and prepare control bits. */
587         skb_reserve(skb, sk->sk_prot->max_header);
588         skb->csum = 0;
589         DCCP_SKB_CB(skb)->dccpd_type = dp->dccps_role == DCCP_ROLE_CLIENT ?
590                                         DCCP_PKT_CLOSE : DCCP_PKT_CLOSEREQ;
591
592         if (active) {
593                 dccp_write_xmit(sk, 1);
594                 dccp_skb_entail(sk, skb);
595                 dccp_transmit_skb(sk, skb_clone(skb, prio));
596                 /* FIXME do we need a retransmit timer here? */
597         } else
598                 dccp_transmit_skb(sk, skb);
599 }