dccp ccid-2: Algorithm to update buffer state
[linux-2.6.git] / net / dccp / ackvec.c
1 /*
2  *  net/dccp/ackvec.c
3  *
4  *  An implementation of Ack Vectors for the DCCP protocol
5  *  Copyright (c) 2007 University of Aberdeen, Scotland, UK
6  *  Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
7  *
8  *      This program is free software; you can redistribute it and/or modify it
9  *      under the terms of the GNU General Public License as published by the
10  *      Free Software Foundation; version 2 of the License;
11  */
12
13 #include "ackvec.h"
14 #include "dccp.h"
15
16 #include <linux/init.h>
17 #include <linux/errno.h>
18 #include <linux/kernel.h>
19 #include <linux/skbuff.h>
20 #include <linux/slab.h>
21
22 #include <net/sock.h>
23
24 static struct kmem_cache *dccp_ackvec_slab;
25 static struct kmem_cache *dccp_ackvec_record_slab;
26
27 struct dccp_ackvec *dccp_ackvec_alloc(const gfp_t priority)
28 {
29         struct dccp_ackvec *av = kmem_cache_zalloc(dccp_ackvec_slab, priority);
30
31         if (av != NULL) {
32                 av->av_buf_head = av->av_buf_tail = DCCPAV_MAX_ACKVEC_LEN - 1;
33                 INIT_LIST_HEAD(&av->av_records);
34         }
35         return av;
36 }
37
38 static void dccp_ackvec_purge_records(struct dccp_ackvec *av)
39 {
40         struct dccp_ackvec_record *cur, *next;
41
42         list_for_each_entry_safe(cur, next, &av->av_records, avr_node)
43                 kmem_cache_free(dccp_ackvec_record_slab, cur);
44         INIT_LIST_HEAD(&av->av_records);
45 }
46
47 void dccp_ackvec_free(struct dccp_ackvec *av)
48 {
49         if (likely(av != NULL)) {
50                 dccp_ackvec_purge_records(av);
51                 kmem_cache_free(dccp_ackvec_slab, av);
52         }
53 }
54
55 /**
56  * dccp_ackvec_update_records  -  Record information about sent Ack Vectors
57  * @av:         Ack Vector records to update
58  * @seqno:      Sequence number of the packet carrying the Ack Vector just sent
59  * @nonce_sum:  The sum of all buffer nonces contained in the Ack Vector
60  */
61 int dccp_ackvec_update_records(struct dccp_ackvec *av, u64 seqno, u8 nonce_sum)
62 {
63         struct dccp_ackvec_record *avr;
64
65         avr = kmem_cache_alloc(dccp_ackvec_record_slab, GFP_ATOMIC);
66         if (avr == NULL)
67                 return -ENOBUFS;
68
69         avr->avr_ack_seqno  = seqno;
70         avr->avr_ack_ptr    = av->av_buf_head;
71         avr->avr_ack_ackno  = av->av_buf_ackno;
72         avr->avr_ack_nonce  = nonce_sum;
73         avr->avr_ack_runlen = dccp_ackvec_runlen(av->av_buf + av->av_buf_head);
74         /*
75          * When the buffer overflows, we keep no more than one record. This is
76          * the simplest way of disambiguating sender-Acks dating from before the
77          * overflow from sender-Acks which refer to after the overflow; a simple
78          * solution is preferable here since we are handling an exception.
79          */
80         if (av->av_overflow)
81                 dccp_ackvec_purge_records(av);
82         /*
83          * Since GSS is incremented for each packet, the list is automatically
84          * arranged in descending order of @ack_seqno.
85          */
86         list_add(&avr->avr_node, &av->av_records);
87
88         dccp_pr_debug("Added Vector, ack_seqno=%llu, ack_ackno=%llu (rl=%u)\n",
89                       (unsigned long long)avr->avr_ack_seqno,
90                       (unsigned long long)avr->avr_ack_ackno,
91                       avr->avr_ack_runlen);
92         return 0;
93 }
94
95 static struct dccp_ackvec_record *dccp_ackvec_lookup(struct list_head *av_list,
96                                                      const u64 ackno)
97 {
98         struct dccp_ackvec_record *avr;
99         /*
100          * Exploit that records are inserted in descending order of sequence
101          * number, start with the oldest record first. If @ackno is `before'
102          * the earliest ack_ackno, the packet is too old to be considered.
103          */
104         list_for_each_entry_reverse(avr, av_list, avr_node) {
105                 if (avr->avr_ack_seqno == ackno)
106                         return avr;
107                 if (before48(ackno, avr->avr_ack_seqno))
108                         break;
109         }
110         return NULL;
111 }
112
113 /*
114  * Buffer index and length computation using modulo-buffersize arithmetic.
115  * Note that, as pointers move from right to left, head is `before' tail.
116  */
117 static inline u16 __ackvec_idx_add(const u16 a, const u16 b)
118 {
119         return (a + b) % DCCPAV_MAX_ACKVEC_LEN;
120 }
121
122 static inline u16 __ackvec_idx_sub(const u16 a, const u16 b)
123 {
124         return __ackvec_idx_add(a, DCCPAV_MAX_ACKVEC_LEN - b);
125 }
126
127 u16 dccp_ackvec_buflen(const struct dccp_ackvec *av)
128 {
129         if (unlikely(av->av_overflow))
130                 return DCCPAV_MAX_ACKVEC_LEN;
131         return __ackvec_idx_sub(av->av_buf_tail, av->av_buf_head);
132 }
133
134 /*
135  * If several packets are missing, the HC-Receiver may prefer to enter multiple
136  * bytes with run length 0, rather than a single byte with a larger run length;
137  * this simplifies table updates if one of the missing packets arrives.
138  */
139 static inline int dccp_ackvec_set_buf_head_state(struct dccp_ackvec *av,
140                                                  const unsigned int packets,
141                                                  const unsigned char state)
142 {
143         unsigned int gap;
144         long new_head;
145
146         if (av->av_vec_len + packets > DCCPAV_MAX_ACKVEC_LEN)
147                 return -ENOBUFS;
148
149         gap      = packets - 1;
150         new_head = av->av_buf_head - packets;
151
152         if (new_head < 0) {
153                 if (gap > 0) {
154                         memset(av->av_buf, DCCPAV_NOT_RECEIVED,
155                                gap + new_head + 1);
156                         gap = -new_head;
157                 }
158                 new_head += DCCPAV_MAX_ACKVEC_LEN;
159         }
160
161         av->av_buf_head = new_head;
162
163         if (gap > 0)
164                 memset(av->av_buf + av->av_buf_head + 1,
165                        DCCPAV_NOT_RECEIVED, gap);
166
167         av->av_buf[av->av_buf_head] = state;
168         av->av_vec_len += packets;
169         return 0;
170 }
171
172 /*
173  * Implements the RFC 4340, Appendix A
174  */
175 int dccp_ackvec_add(struct dccp_ackvec *av, const struct sock *sk,
176                     const u64 ackno, const u8 state)
177 {
178         u8 *cur_head = av->av_buf + av->av_buf_head,
179            *buf_end  = av->av_buf + DCCPAV_MAX_ACKVEC_LEN;
180         /*
181          * Check at the right places if the buffer is full, if it is, tell the
182          * caller to start dropping packets till the HC-Sender acks our ACK
183          * vectors, when we will free up space in av_buf.
184          *
185          * We may well decide to do buffer compression, etc, but for now lets
186          * just drop.
187          *
188          * From Appendix A.1.1 (`New Packets'):
189          *
190          *      Of course, the circular buffer may overflow, either when the
191          *      HC-Sender is sending data at a very high rate, when the
192          *      HC-Receiver's acknowledgements are not reaching the HC-Sender,
193          *      or when the HC-Sender is forgetting to acknowledge those acks
194          *      (so the HC-Receiver is unable to clean up old state). In this
195          *      case, the HC-Receiver should either compress the buffer (by
196          *      increasing run lengths when possible), transfer its state to
197          *      a larger buffer, or, as a last resort, drop all received
198          *      packets, without processing them whatsoever, until its buffer
199          *      shrinks again.
200          */
201
202         /* See if this is the first ackno being inserted */
203         if (av->av_vec_len == 0) {
204                 *cur_head = state;
205                 av->av_vec_len = 1;
206         } else if (after48(ackno, av->av_buf_ackno)) {
207                 const u64 delta = dccp_delta_seqno(av->av_buf_ackno, ackno);
208
209                 /*
210                  * Look if the state of this packet is the same as the
211                  * previous ackno and if so if we can bump the head len.
212                  */
213                 if (delta == 1 && dccp_ackvec_state(cur_head) == state &&
214                     dccp_ackvec_runlen(cur_head) < DCCPAV_MAX_RUNLEN)
215                         *cur_head += 1;
216                 else if (dccp_ackvec_set_buf_head_state(av, delta, state))
217                         return -ENOBUFS;
218         } else {
219                 /*
220                  * A.1.2.  Old Packets
221                  *
222                  *      When a packet with Sequence Number S <= buf_ackno
223                  *      arrives, the HC-Receiver will scan the table for
224                  *      the byte corresponding to S. (Indexing structures
225                  *      could reduce the complexity of this scan.)
226                  */
227                 u64 delta = dccp_delta_seqno(ackno, av->av_buf_ackno);
228
229                 while (1) {
230                         const u8 len = dccp_ackvec_runlen(cur_head);
231                         /*
232                          * valid packets not yet in av_buf have a reserved
233                          * entry, with a len equal to 0.
234                          */
235                         if (*cur_head == DCCPAV_NOT_RECEIVED && delta == 0) {
236                                 dccp_pr_debug("Found %llu reserved seat!\n",
237                                               (unsigned long long)ackno);
238                                 *cur_head = state;
239                                 goto out;
240                         }
241                         /* len == 0 means one packet */
242                         if (delta < len + 1)
243                                 goto out_duplicate;
244
245                         delta -= len + 1;
246                         if (++cur_head == buf_end)
247                                 cur_head = av->av_buf;
248                 }
249         }
250
251         av->av_buf_ackno = ackno;
252 out:
253         return 0;
254
255 out_duplicate:
256         /* Duplicate packet */
257         dccp_pr_debug("Received a dup or already considered lost "
258                       "packet: %llu\n", (unsigned long long)ackno);
259         return -EILSEQ;
260 }
261
262 static void dccp_ackvec_throw_record(struct dccp_ackvec *av,
263                                      struct dccp_ackvec_record *avr)
264 {
265         struct dccp_ackvec_record *next;
266
267         /* sort out vector length */
268         if (av->av_buf_head <= avr->avr_ack_ptr)
269                 av->av_vec_len = avr->avr_ack_ptr - av->av_buf_head;
270         else
271                 av->av_vec_len = DCCPAV_MAX_ACKVEC_LEN - 1 -
272                                  av->av_buf_head + avr->avr_ack_ptr;
273
274         /* free records */
275         list_for_each_entry_safe_from(avr, next, &av->av_records, avr_node) {
276                 list_del(&avr->avr_node);
277                 kmem_cache_free(dccp_ackvec_record_slab, avr);
278         }
279 }
280
281 void dccp_ackvec_check_rcv_ackno(struct dccp_ackvec *av, struct sock *sk,
282                                  const u64 ackno)
283 {
284         struct dccp_ackvec_record *avr;
285
286         /*
287          * If we traverse backwards, it should be faster when we have large
288          * windows. We will be receiving ACKs for stuff we sent a while back
289          * -sorbo.
290          */
291         list_for_each_entry_reverse(avr, &av->av_records, avr_node) {
292                 if (ackno == avr->avr_ack_seqno) {
293                         dccp_pr_debug("%s ACK packet 0, len=%d, ack_seqno=%llu, "
294                                       "ack_ackno=%llu, ACKED!\n",
295                                       dccp_role(sk), avr->avr_ack_runlen,
296                                       (unsigned long long)avr->avr_ack_seqno,
297                                       (unsigned long long)avr->avr_ack_ackno);
298                         dccp_ackvec_throw_record(av, avr);
299                         break;
300                 } else if (avr->avr_ack_seqno > ackno)
301                         break; /* old news */
302         }
303 }
304
305 static void dccp_ackvec_check_rcv_ackvector(struct dccp_ackvec *av,
306                                             struct sock *sk, u64 *ackno,
307                                             const unsigned char len,
308                                             const unsigned char *vector)
309 {
310         unsigned char i;
311         struct dccp_ackvec_record *avr;
312
313         /* Check if we actually sent an ACK vector */
314         if (list_empty(&av->av_records))
315                 return;
316
317         i = len;
318         /*
319          * XXX
320          * I think it might be more efficient to work backwards. See comment on
321          * rcv_ackno. -sorbo.
322          */
323         avr = list_entry(av->av_records.next, struct dccp_ackvec_record, avr_node);
324         while (i--) {
325                 const u8 rl = dccp_ackvec_runlen(vector);
326                 u64 ackno_end_rl;
327
328                 dccp_set_seqno(&ackno_end_rl, *ackno - rl);
329
330                 /*
331                  * If our AVR sequence number is greater than the ack, go
332                  * forward in the AVR list until it is not so.
333                  */
334                 list_for_each_entry_from(avr, &av->av_records, avr_node) {
335                         if (!after48(avr->avr_ack_seqno, *ackno))
336                                 goto found;
337                 }
338                 /* End of the av_records list, not found, exit */
339                 break;
340 found:
341                 if (between48(avr->avr_ack_seqno, ackno_end_rl, *ackno)) {
342                         if (dccp_ackvec_state(vector) != DCCPAV_NOT_RECEIVED) {
343                                 dccp_pr_debug("%s ACK vector 0, len=%d, "
344                                               "ack_seqno=%llu, ack_ackno=%llu, "
345                                               "ACKED!\n",
346                                               dccp_role(sk), len,
347                                               (unsigned long long)
348                                               avr->avr_ack_seqno,
349                                               (unsigned long long)
350                                               avr->avr_ack_ackno);
351                                 dccp_ackvec_throw_record(av, avr);
352                                 break;
353                         }
354                         /*
355                          * If it wasn't received, continue scanning... we might
356                          * find another one.
357                          */
358                 }
359
360                 dccp_set_seqno(ackno, ackno_end_rl - 1);
361                 ++vector;
362         }
363 }
364
365 int dccp_ackvec_parse(struct sock *sk, const struct sk_buff *skb,
366                       u64 *ackno, const u8 opt, const u8 *value, const u8 len)
367 {
368         if (len > DCCP_SINGLE_OPT_MAXLEN)
369                 return -1;
370
371         /* dccp_ackvector_print(DCCP_SKB_CB(skb)->dccpd_ack_seq, value, len); */
372         dccp_ackvec_check_rcv_ackvector(dccp_sk(sk)->dccps_hc_rx_ackvec, sk,
373                                         ackno, len, value);
374         return 0;
375 }
376
377 /**
378  * dccp_ackvec_clear_state  -  Perform house-keeping / garbage-collection
379  * This routine is called when the peer acknowledges the receipt of Ack Vectors
380  * up to and including @ackno. While based on on section A.3 of RFC 4340, here
381  * are additional precautions to prevent corrupted buffer state. In particular,
382  * we use tail_ackno to identify outdated records; it always marks the earliest
383  * packet of group (2) in 11.4.2.
384  */
385 void dccp_ackvec_clear_state(struct dccp_ackvec *av, const u64 ackno)
386  {
387         struct dccp_ackvec_record *avr, *next;
388         u8 runlen_now, eff_runlen;
389         s64 delta;
390
391         avr = dccp_ackvec_lookup(&av->av_records, ackno);
392         if (avr == NULL)
393                 return;
394         /*
395          * Deal with outdated acknowledgments: this arises when e.g. there are
396          * several old records and the acks from the peer come in slowly. In
397          * that case we may still have records that pre-date tail_ackno.
398          */
399         delta = dccp_delta_seqno(av->av_tail_ackno, avr->avr_ack_ackno);
400         if (delta < 0)
401                 goto free_records;
402         /*
403          * Deal with overlapping Ack Vectors: don't subtract more than the
404          * number of packets between tail_ackno and ack_ackno.
405          */
406         eff_runlen = delta < avr->avr_ack_runlen ? delta : avr->avr_ack_runlen;
407
408         runlen_now = dccp_ackvec_runlen(av->av_buf + avr->avr_ack_ptr);
409         /*
410          * The run length of Ack Vector cells does not decrease over time. If
411          * the run length is the same as at the time the Ack Vector was sent, we
412          * free the ack_ptr cell. That cell can however not be freed if the run
413          * length has increased: in this case we need to move the tail pointer
414          * backwards (towards higher indices), to its next-oldest neighbour.
415          */
416         if (runlen_now > eff_runlen) {
417
418                 av->av_buf[avr->avr_ack_ptr] -= eff_runlen + 1;
419                 av->av_buf_tail = __ackvec_idx_add(avr->avr_ack_ptr, 1);
420
421                 /* This move may not have cleared the overflow flag. */
422                 if (av->av_overflow)
423                         av->av_overflow = (av->av_buf_head == av->av_buf_tail);
424         } else {
425                 av->av_buf_tail = avr->avr_ack_ptr;
426                 /*
427                  * We have made sure that avr points to a valid cell within the
428                  * buffer. This cell is either older than head, or equals head
429                  * (empty buffer): in both cases we no longer have any overflow.
430                  */
431                 av->av_overflow = 0;
432         }
433
434         /*
435          * The peer has acknowledged up to and including ack_ackno. Hence the
436          * first packet in group (2) of 11.4.2 is the successor of ack_ackno.
437          */
438         av->av_tail_ackno = ADD48(avr->avr_ack_ackno, 1);
439
440 free_records:
441         list_for_each_entry_safe_from(avr, next, &av->av_records, avr_node) {
442                 list_del(&avr->avr_node);
443                 kmem_cache_free(dccp_ackvec_record_slab, avr);
444         }
445 }
446
447 int __init dccp_ackvec_init(void)
448 {
449         dccp_ackvec_slab = kmem_cache_create("dccp_ackvec",
450                                              sizeof(struct dccp_ackvec), 0,
451                                              SLAB_HWCACHE_ALIGN, NULL);
452         if (dccp_ackvec_slab == NULL)
453                 goto out_err;
454
455         dccp_ackvec_record_slab = kmem_cache_create("dccp_ackvec_record",
456                                              sizeof(struct dccp_ackvec_record),
457                                              0, SLAB_HWCACHE_ALIGN, NULL);
458         if (dccp_ackvec_record_slab == NULL)
459                 goto out_destroy_slab;
460
461         return 0;
462
463 out_destroy_slab:
464         kmem_cache_destroy(dccp_ackvec_slab);
465         dccp_ackvec_slab = NULL;
466 out_err:
467         DCCP_CRIT("Unable to create Ack Vector slab cache");
468         return -ENOBUFS;
469 }
470
471 void dccp_ackvec_exit(void)
472 {
473         if (dccp_ackvec_slab != NULL) {
474                 kmem_cache_destroy(dccp_ackvec_slab);
475                 dccp_ackvec_slab = NULL;
476         }
477         if (dccp_ackvec_record_slab != NULL) {
478                 kmem_cache_destroy(dccp_ackvec_record_slab);
479                 dccp_ackvec_record_slab = NULL;
480         }
481 }