This commit was manufactured by cvs2svn to create branch 'vserver'.
[linux-2.6.git] / net / dccp / options.c
1 /*
2  *  net/dccp/options.c
3  *
4  *  An implementation of the DCCP protocol
5  *  Copyright (c) 2005 Aristeu Sergio Rozanski Filho <aris@cathedrallabs.org>
6  *  Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
7  *  Copyright (c) 2005 Ian McDonald <iam4@cs.waikato.ac.nz>
8  *
9  *      This program is free software; you can redistribute it and/or
10  *      modify it under the terms of the GNU General Public License
11  *      as published by the Free Software Foundation; either version
12  *      2 of the License, or (at your option) any later version.
13  */
14 #include <linux/config.h>
15 #include <linux/dccp.h>
16 #include <linux/module.h>
17 #include <linux/types.h>
18 #include <linux/kernel.h>
19 #include <linux/skbuff.h>
20
21 #include "ackvec.h"
22 #include "ccid.h"
23 #include "dccp.h"
24 #include "feat.h"
25
26 int dccp_feat_default_sequence_window = DCCPF_INITIAL_SEQUENCE_WINDOW;
27 int dccp_feat_default_rx_ccid         = DCCPF_INITIAL_CCID;
28 int dccp_feat_default_tx_ccid         = DCCPF_INITIAL_CCID;
29 int dccp_feat_default_ack_ratio       = DCCPF_INITIAL_ACK_RATIO;
30 int dccp_feat_default_send_ack_vector = DCCPF_INITIAL_SEND_ACK_VECTOR;
31 int dccp_feat_default_send_ndp_count  = DCCPF_INITIAL_SEND_NDP_COUNT;
32
33 void dccp_minisock_init(struct dccp_minisock *dmsk)
34 {
35         dmsk->dccpms_sequence_window = dccp_feat_default_sequence_window;
36         dmsk->dccpms_rx_ccid         = dccp_feat_default_rx_ccid;
37         dmsk->dccpms_tx_ccid         = dccp_feat_default_tx_ccid;
38         dmsk->dccpms_ack_ratio       = dccp_feat_default_ack_ratio;
39         dmsk->dccpms_send_ack_vector = dccp_feat_default_send_ack_vector;
40         dmsk->dccpms_send_ndp_count  = dccp_feat_default_send_ndp_count;
41 }
42
43 static u32 dccp_decode_value_var(const unsigned char *bf, const u8 len)
44 {
45         u32 value = 0;
46
47         if (len > 3)
48                 value += *bf++ << 24;
49         if (len > 2)
50                 value += *bf++ << 16;
51         if (len > 1)
52                 value += *bf++ << 8;
53         if (len > 0)
54                 value += *bf;
55
56         return value;
57 }
58
59 int dccp_parse_options(struct sock *sk, struct sk_buff *skb)
60 {
61         struct dccp_sock *dp = dccp_sk(sk);
62 #ifdef CONFIG_IP_DCCP_DEBUG
63         const char *debug_prefix = dp->dccps_role == DCCP_ROLE_CLIENT ?
64                                         "CLIENT rx opt: " : "server rx opt: ";
65 #endif
66         const struct dccp_hdr *dh = dccp_hdr(skb);
67         const u8 pkt_type = DCCP_SKB_CB(skb)->dccpd_type;
68         unsigned char *options = (unsigned char *)dh + dccp_hdr_len(skb);
69         unsigned char *opt_ptr = options;
70         const unsigned char *opt_end = (unsigned char *)dh +
71                                         (dh->dccph_doff * 4);
72         struct dccp_options_received *opt_recv = &dp->dccps_options_received;
73         unsigned char opt, len;
74         unsigned char *value;
75         u32 elapsed_time;
76         int rc;
77         int mandatory = 0;
78
79         memset(opt_recv, 0, sizeof(*opt_recv));
80
81         opt = len = 0;
82         while (opt_ptr != opt_end) {
83                 opt   = *opt_ptr++;
84                 len   = 0;
85                 value = NULL;
86
87                 /* Check if this isn't a single byte option */
88                 if (opt > DCCPO_MAX_RESERVED) {
89                         if (opt_ptr == opt_end)
90                                 goto out_invalid_option;
91
92                         len = *opt_ptr++;
93                         if (len < 3)
94                                 goto out_invalid_option;
95                         /*
96                          * Remove the type and len fields, leaving
97                          * just the value size
98                          */
99                         len     -= 2;
100                         value   = opt_ptr;
101                         opt_ptr += len;
102
103                         if (opt_ptr > opt_end)
104                                 goto out_invalid_option;
105                 }
106
107                 switch (opt) {
108                 case DCCPO_PADDING:
109                         break;
110                 case DCCPO_MANDATORY:
111                         if (mandatory)
112                                 goto out_invalid_option;
113                         if (pkt_type != DCCP_PKT_DATA)
114                                 mandatory = 1;
115                         break;
116                 case DCCPO_NDP_COUNT:
117                         if (len > 3)
118                                 goto out_invalid_option;
119
120                         opt_recv->dccpor_ndp = dccp_decode_value_var(value, len);
121                         dccp_pr_debug("%sNDP count=%d\n", debug_prefix,
122                                       opt_recv->dccpor_ndp);
123                         break;
124                 case DCCPO_CHANGE_L:
125                         /* fall through */
126                 case DCCPO_CHANGE_R:
127                         if (len < 2)
128                                 goto out_invalid_option;
129                         rc = dccp_feat_change_recv(sk, opt, *value, value + 1,
130                                                    len - 1);
131                         /*
132                          * When there is a change error, change_recv is
133                          * responsible for dealing with it.  i.e. reply with an
134                          * empty confirm.
135                          * If the change was mandatory, then we need to die.
136                          */
137                         if (rc && mandatory)
138                                 goto out_invalid_option;
139                         break;
140                 case DCCPO_CONFIRM_L:
141                         /* fall through */
142                 case DCCPO_CONFIRM_R:
143                         if (len < 2)
144                                 goto out_invalid_option;
145                         if (dccp_feat_confirm_recv(sk, opt, *value,
146                                                    value + 1, len - 1))
147                                 goto out_invalid_option;
148                         break;
149                 case DCCPO_ACK_VECTOR_0:
150                 case DCCPO_ACK_VECTOR_1:
151                         if (pkt_type == DCCP_PKT_DATA)
152                                 break;
153
154                         if (dccp_msk(sk)->dccpms_send_ack_vector &&
155                             dccp_ackvec_parse(sk, skb, opt, value, len))
156                                 goto out_invalid_option;
157                         break;
158                 case DCCPO_TIMESTAMP:
159                         if (len != 4)
160                                 goto out_invalid_option;
161
162                         opt_recv->dccpor_timestamp = ntohl(*(__be32 *)value);
163
164                         dp->dccps_timestamp_echo = opt_recv->dccpor_timestamp;
165                         dccp_timestamp(sk, &dp->dccps_timestamp_time);
166
167                         dccp_pr_debug("%sTIMESTAMP=%u, ackno=%llu\n",
168                                       debug_prefix, opt_recv->dccpor_timestamp,
169                                       (unsigned long long)
170                                       DCCP_SKB_CB(skb)->dccpd_ack_seq);
171                         break;
172                 case DCCPO_TIMESTAMP_ECHO:
173                         if (len != 4 && len != 6 && len != 8)
174                                 goto out_invalid_option;
175
176                         opt_recv->dccpor_timestamp_echo = ntohl(*(__be32 *)value);
177
178                         dccp_pr_debug("%sTIMESTAMP_ECHO=%u, len=%d, ackno=%llu, ",
179                                       debug_prefix,
180                                       opt_recv->dccpor_timestamp_echo,
181                                       len + 2,
182                                       (unsigned long long)
183                                       DCCP_SKB_CB(skb)->dccpd_ack_seq);
184
185
186                         if (len == 4)
187                                 break;
188
189                         if (len == 6)
190                                 elapsed_time = ntohs(*(__be16 *)(value + 4));
191                         else
192                                 elapsed_time = ntohl(*(__be32 *)(value + 4));
193
194                         /* Give precedence to the biggest ELAPSED_TIME */
195                         if (elapsed_time > opt_recv->dccpor_elapsed_time)
196                                 opt_recv->dccpor_elapsed_time = elapsed_time;
197                         break;
198                 case DCCPO_ELAPSED_TIME:
199                         if (len != 2 && len != 4)
200                                 goto out_invalid_option;
201
202                         if (pkt_type == DCCP_PKT_DATA)
203                                 continue;
204
205                         if (len == 2)
206                                 elapsed_time = ntohs(*(__be16 *)value);
207                         else
208                                 elapsed_time = ntohl(*(__be32 *)value);
209
210                         if (elapsed_time > opt_recv->dccpor_elapsed_time)
211                                 opt_recv->dccpor_elapsed_time = elapsed_time;
212
213                         dccp_pr_debug("%sELAPSED_TIME=%d\n", debug_prefix,
214                                       elapsed_time);
215                         break;
216                         /*
217                          * From draft-ietf-dccp-spec-11.txt:
218                          *
219                          *      Option numbers 128 through 191 are for
220                          *      options sent from the HC-Sender to the
221                          *      HC-Receiver; option numbers 192 through 255
222                          *      are for options sent from the HC-Receiver to
223                          *      the HC-Sender.
224                          */
225                 case 128 ... 191: {
226                         const u16 idx = value - options;
227
228                         if (ccid_hc_rx_parse_options(dp->dccps_hc_rx_ccid, sk,
229                                                      opt, len, idx,
230                                                      value) != 0)
231                                 goto out_invalid_option;
232                 }
233                         break;
234                 case 192 ... 255: {
235                         const u16 idx = value - options;
236
237                         if (ccid_hc_tx_parse_options(dp->dccps_hc_tx_ccid, sk,
238                                                      opt, len, idx,
239                                                      value) != 0)
240                                 goto out_invalid_option;
241                 }
242                         break;
243                 default:
244                         pr_info("DCCP(%p): option %d(len=%d) not "
245                                 "implemented, ignoring\n",
246                                 sk, opt, len);
247                         break;
248                 }
249
250                 if (opt != DCCPO_MANDATORY)
251                         mandatory = 0;
252         }
253
254         /* mandatory was the last byte in option list -> reset connection */
255         if (mandatory)
256                 goto out_invalid_option;
257
258         return 0;
259
260 out_invalid_option:
261         DCCP_INC_STATS_BH(DCCP_MIB_INVALIDOPT);
262         DCCP_SKB_CB(skb)->dccpd_reset_code = DCCP_RESET_CODE_OPTION_ERROR;
263         pr_info("DCCP(%p): invalid option %d, len=%d\n", sk, opt, len);
264         return -1;
265 }
266
267 EXPORT_SYMBOL_GPL(dccp_parse_options);
268
269 static void dccp_encode_value_var(const u32 value, unsigned char *to,
270                                   const unsigned int len)
271 {
272         if (len > 3)
273                 *to++ = (value & 0xFF000000) >> 24;
274         if (len > 2)
275                 *to++ = (value & 0xFF0000) >> 16;
276         if (len > 1)
277                 *to++ = (value & 0xFF00) >> 8;
278         if (len > 0)
279                 *to++ = (value & 0xFF);
280 }
281
282 static inline int dccp_ndp_len(const int ndp)
283 {
284         return likely(ndp <= 0xFF) ? 1 : ndp <= 0xFFFF ? 2 : 3;
285 }
286
287 int dccp_insert_option(struct sock *sk, struct sk_buff *skb,
288                         const unsigned char option,
289                         const void *value, const unsigned char len)
290 {
291         unsigned char *to;
292
293         if (DCCP_SKB_CB(skb)->dccpd_opt_len + len + 2 > DCCP_MAX_OPT_LEN)
294                 return -1;
295
296         DCCP_SKB_CB(skb)->dccpd_opt_len += len + 2;
297
298         to    = skb_push(skb, len + 2);
299         *to++ = option;
300         *to++ = len + 2;
301
302         memcpy(to, value, len);
303         return 0;
304 }
305
306 EXPORT_SYMBOL_GPL(dccp_insert_option);
307
308 static int dccp_insert_option_ndp(struct sock *sk, struct sk_buff *skb)
309 {
310         struct dccp_sock *dp = dccp_sk(sk);
311         int ndp = dp->dccps_ndp_count;
312
313         if (dccp_non_data_packet(skb))
314                 ++dp->dccps_ndp_count;
315         else
316                 dp->dccps_ndp_count = 0;
317
318         if (ndp > 0) {
319                 unsigned char *ptr;
320                 const int ndp_len = dccp_ndp_len(ndp);
321                 const int len = ndp_len + 2;
322
323                 if (DCCP_SKB_CB(skb)->dccpd_opt_len + len > DCCP_MAX_OPT_LEN)
324                         return -1;
325
326                 DCCP_SKB_CB(skb)->dccpd_opt_len += len;
327
328                 ptr = skb_push(skb, len);
329                 *ptr++ = DCCPO_NDP_COUNT;
330                 *ptr++ = len;
331                 dccp_encode_value_var(ndp, ptr, ndp_len);
332         }
333
334         return 0;
335 }
336
337 static inline int dccp_elapsed_time_len(const u32 elapsed_time)
338 {
339         return elapsed_time == 0 ? 0 : elapsed_time <= 0xFFFF ? 2 : 4;
340 }
341
342 int dccp_insert_option_elapsed_time(struct sock *sk, struct sk_buff *skb,
343                                     u32 elapsed_time)
344 {
345         const int elapsed_time_len = dccp_elapsed_time_len(elapsed_time);
346         const int len = 2 + elapsed_time_len;
347         unsigned char *to;
348
349         if (elapsed_time_len == 0)
350                 return 0;
351
352         if (DCCP_SKB_CB(skb)->dccpd_opt_len + len > DCCP_MAX_OPT_LEN)
353                 return -1;
354
355         DCCP_SKB_CB(skb)->dccpd_opt_len += len;
356
357         to    = skb_push(skb, len);
358         *to++ = DCCPO_ELAPSED_TIME;
359         *to++ = len;
360
361         if (elapsed_time_len == 2) {
362                 const __be16 var16 = htons((u16)elapsed_time);
363                 memcpy(to, &var16, 2);
364         } else {
365                 const __be32 var32 = htonl(elapsed_time);
366                 memcpy(to, &var32, 4);
367         }
368
369         return 0;
370 }
371
372 EXPORT_SYMBOL_GPL(dccp_insert_option_elapsed_time);
373
374 void dccp_timestamp(const struct sock *sk, struct timeval *tv)
375 {
376         const struct dccp_sock *dp = dccp_sk(sk);
377
378         do_gettimeofday(tv);
379         tv->tv_sec  -= dp->dccps_epoch.tv_sec;
380         tv->tv_usec -= dp->dccps_epoch.tv_usec;
381
382         while (tv->tv_usec < 0) {
383                 tv->tv_sec--;
384                 tv->tv_usec += USEC_PER_SEC;
385         }
386 }
387
388 EXPORT_SYMBOL_GPL(dccp_timestamp);
389
390 int dccp_insert_option_timestamp(struct sock *sk, struct sk_buff *skb)
391 {
392         struct timeval tv;
393         __be32 now;
394
395         dccp_timestamp(sk, &tv);
396         now = htonl(timeval_usecs(&tv) / 10);
397         /* yes this will overflow but that is the point as we want a
398          * 10 usec 32 bit timer which mean it wraps every 11.9 hours */
399
400         return dccp_insert_option(sk, skb, DCCPO_TIMESTAMP, &now, sizeof(now));
401 }
402
403 EXPORT_SYMBOL_GPL(dccp_insert_option_timestamp);
404
405 static int dccp_insert_option_timestamp_echo(struct sock *sk,
406                                              struct sk_buff *skb)
407 {
408         struct dccp_sock *dp = dccp_sk(sk);
409         struct timeval now;
410         __be32 tstamp_echo;
411         u32 elapsed_time;
412         int len, elapsed_time_len;
413         unsigned char *to;
414
415         dccp_timestamp(sk, &now);
416         elapsed_time = timeval_delta(&now, &dp->dccps_timestamp_time) / 10;
417         elapsed_time_len = dccp_elapsed_time_len(elapsed_time);
418         len = 6 + elapsed_time_len;
419
420         if (DCCP_SKB_CB(skb)->dccpd_opt_len + len > DCCP_MAX_OPT_LEN)
421                 return -1;
422
423         DCCP_SKB_CB(skb)->dccpd_opt_len += len;
424
425         to    = skb_push(skb, len);
426         *to++ = DCCPO_TIMESTAMP_ECHO;
427         *to++ = len;
428
429         tstamp_echo = htonl(dp->dccps_timestamp_echo);
430         memcpy(to, &tstamp_echo, 4);
431         to += 4;
432
433         if (elapsed_time_len == 2) {
434                 const __be16 var16 = htons((u16)elapsed_time);
435                 memcpy(to, &var16, 2);
436         } else if (elapsed_time_len == 4) {
437                 const __be32 var32 = htonl(elapsed_time);
438                 memcpy(to, &var32, 4);
439         }
440
441         dp->dccps_timestamp_echo = 0;
442         dp->dccps_timestamp_time.tv_sec = 0;
443         dp->dccps_timestamp_time.tv_usec = 0;
444         return 0;
445 }
446
447 static int dccp_insert_feat_opt(struct sk_buff *skb, u8 type, u8 feat,
448                                 u8 *val, u8 len)
449 {
450         u8 *to;
451
452         if (DCCP_SKB_CB(skb)->dccpd_opt_len + len + 3 > DCCP_MAX_OPT_LEN) {
453                 LIMIT_NETDEBUG(KERN_INFO "DCCP: packet too small"
454                                " to insert feature %d option!\n", feat);
455                 return -1;
456         }
457
458         DCCP_SKB_CB(skb)->dccpd_opt_len += len + 3;
459
460         to    = skb_push(skb, len + 3);
461         *to++ = type;
462         *to++ = len + 3;
463         *to++ = feat;
464
465         if (len)
466                 memcpy(to, val, len);
467         dccp_pr_debug("option %d feat %d len %d\n", type, feat, len);
468
469         return 0;
470 }
471
472 static int dccp_insert_options_feat(struct sock *sk, struct sk_buff *skb)
473 {
474         struct dccp_sock *dp = dccp_sk(sk);
475         struct dccp_minisock *dmsk = dccp_msk(sk);
476         struct dccp_opt_pend *opt, *next;
477         int change = 0;
478
479         /* confirm any options [NN opts] */
480         list_for_each_entry_safe(opt, next, &dmsk->dccpms_conf, dccpop_node) {
481                 dccp_insert_feat_opt(skb, opt->dccpop_type,
482                                      opt->dccpop_feat, opt->dccpop_val,
483                                      opt->dccpop_len);
484                 /* fear empty confirms */
485                 if (opt->dccpop_val)
486                         kfree(opt->dccpop_val);
487                 kfree(opt);
488         }
489         INIT_LIST_HEAD(&dmsk->dccpms_conf);
490
491         /* see which features we need to send */
492         list_for_each_entry(opt, &dmsk->dccpms_pending, dccpop_node) {
493                 /* see if we need to send any confirm */
494                 if (opt->dccpop_sc) {
495                         dccp_insert_feat_opt(skb, opt->dccpop_type + 1,
496                                              opt->dccpop_feat,
497                                              opt->dccpop_sc->dccpoc_val,
498                                              opt->dccpop_sc->dccpoc_len);
499
500                         BUG_ON(!opt->dccpop_sc->dccpoc_val);
501                         kfree(opt->dccpop_sc->dccpoc_val);
502                         kfree(opt->dccpop_sc);
503                         opt->dccpop_sc = NULL;
504                 }
505
506                 /* any option not confirmed, re-send it */
507                 if (!opt->dccpop_conf) {
508                         dccp_insert_feat_opt(skb, opt->dccpop_type,
509                                              opt->dccpop_feat, opt->dccpop_val,
510                                              opt->dccpop_len);
511                         change++;
512                 }
513         }
514
515         /* Retransmit timer.
516          * If this is the master listening sock, we don't set a timer on it.  It
517          * should be fine because if the dude doesn't receive our RESPONSE
518          * [which will contain the CHANGE] he will send another REQUEST which
519          * will "retrnasmit" the change.
520          */
521         if (change && dp->dccps_role != DCCP_ROLE_LISTEN) {
522                 dccp_pr_debug("reset feat negotiation timer %p\n", sk);
523
524                 /* XXX don't reset the timer on re-transmissions.  I.e. reset it
525                  * only when sending new stuff i guess.  Currently the timer
526                  * never backs off because on re-transmission it just resets it!
527                  */
528                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
529                                           inet_csk(sk)->icsk_rto, DCCP_RTO_MAX);
530         }
531
532         return 0;
533 }
534
535 int dccp_insert_options(struct sock *sk, struct sk_buff *skb)
536 {
537         struct dccp_sock *dp = dccp_sk(sk);
538         struct dccp_minisock *dmsk = dccp_msk(sk);
539
540         DCCP_SKB_CB(skb)->dccpd_opt_len = 0;
541
542         if (dmsk->dccpms_send_ndp_count &&
543             dccp_insert_option_ndp(sk, skb))
544                 return -1;
545
546         if (!dccp_packet_without_ack(skb)) {
547                 if (dmsk->dccpms_send_ack_vector &&
548                     dccp_ackvec_pending(dp->dccps_hc_rx_ackvec) &&
549                     dccp_insert_option_ackvec(sk, skb))
550                         return -1;
551
552                 if (dp->dccps_timestamp_echo != 0 &&
553                     dccp_insert_option_timestamp_echo(sk, skb))
554                         return -1;
555         }
556
557         if (dp->dccps_hc_rx_insert_options) {
558                 if (ccid_hc_rx_insert_options(dp->dccps_hc_rx_ccid, sk, skb))
559                         return -1;
560                 dp->dccps_hc_rx_insert_options = 0;
561         }
562         if (dp->dccps_hc_tx_insert_options) {
563                 if (ccid_hc_tx_insert_options(dp->dccps_hc_tx_ccid, sk, skb))
564                         return -1;
565                 dp->dccps_hc_tx_insert_options = 0;
566         }
567
568         /* Feature negotiation */
569         /* Data packets can't do feat negotiation */
570         if (DCCP_SKB_CB(skb)->dccpd_type != DCCP_PKT_DATA &&
571             DCCP_SKB_CB(skb)->dccpd_type != DCCP_PKT_DATAACK &&
572             dccp_insert_options_feat(sk, skb))
573                 return -1;
574
575         /* XXX: insert other options when appropriate */
576
577         if (DCCP_SKB_CB(skb)->dccpd_opt_len != 0) {
578                 /* The length of all options has to be a multiple of 4 */
579                 int padding = DCCP_SKB_CB(skb)->dccpd_opt_len % 4;
580
581                 if (padding != 0) {
582                         padding = 4 - padding;
583                         memset(skb_push(skb, padding), 0, padding);
584                         DCCP_SKB_CB(skb)->dccpd_opt_len += padding;
585                 }
586         }
587
588         return 0;
589 }