[DCCP] ACKVEC: Optimization - Do not traverse records if none will be found
[linux-2.6.git] / net / dccp / ackvec.c
1 /*
2  *  net/dccp/ackvec.c
3  *
4  *  An implementation of the DCCP protocol
5  *  Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
6  *
7  *      This program is free software; you can redistribute it and/or modify it
8  *      under the terms of the GNU General Public License as published by the
9  *      Free Software Foundation; version 2 of the License;
10  */
11
12 #include "ackvec.h"
13 #include "dccp.h"
14
15 #include <linux/dccp.h>
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 kmem_cache_t *dccp_ackvec_slab;
25 static kmem_cache_t *dccp_ackvec_record_slab;
26
27 static struct dccp_ackvec_record *dccp_ackvec_record_new(void)
28 {
29         struct dccp_ackvec_record *avr =
30                         kmem_cache_alloc(dccp_ackvec_record_slab, GFP_ATOMIC);
31
32         if (avr != NULL)
33                 INIT_LIST_HEAD(&avr->dccpavr_node);
34
35         return avr;
36 }
37
38 static void dccp_ackvec_record_delete(struct dccp_ackvec_record *avr)
39 {
40         if (unlikely(avr == NULL))
41                 return;
42         /* Check if deleting a linked record */
43         WARN_ON(!list_empty(&avr->dccpavr_node));
44         kmem_cache_free(dccp_ackvec_record_slab, avr);
45 }
46
47 static void dccp_ackvec_insert_avr(struct dccp_ackvec *av,
48                                    struct dccp_ackvec_record *avr)
49 {
50         /*
51          * AVRs are sorted by seqno. Since we are sending them in order, we
52          * just add the AVR at the head of the list.
53          * -sorbo.
54          */
55         if (!list_empty(&av->dccpav_records)) {
56                 const struct dccp_ackvec_record *head =
57                                         list_entry(av->dccpav_records.next,
58                                                    struct dccp_ackvec_record,
59                                                    dccpavr_node);
60                 BUG_ON(before48(avr->dccpavr_ack_seqno,
61                                 head->dccpavr_ack_seqno));
62         }
63
64         list_add(&avr->dccpavr_node, &av->dccpav_records);
65 }
66
67 int dccp_insert_option_ackvec(struct sock *sk, struct sk_buff *skb)
68 {
69         struct dccp_sock *dp = dccp_sk(sk);
70         struct dccp_ackvec *av = dp->dccps_hc_rx_ackvec;
71         int len = av->dccpav_vec_len + 2;
72         struct timeval now;
73         u32 elapsed_time;
74         unsigned char *to, *from;
75         struct dccp_ackvec_record *avr;
76
77         if (DCCP_SKB_CB(skb)->dccpd_opt_len + len > DCCP_MAX_OPT_LEN)
78                 return -1;
79
80         dccp_timestamp(sk, &now);
81         elapsed_time = timeval_delta(&now, &av->dccpav_time) / 10;
82
83         if (elapsed_time != 0 &&
84             dccp_insert_option_elapsed_time(sk, skb, elapsed_time))
85                 return -1;
86
87         avr = dccp_ackvec_record_new();
88         if (avr == NULL)
89                 return -1;
90
91         DCCP_SKB_CB(skb)->dccpd_opt_len += len;
92
93         to    = skb_push(skb, len);
94         *to++ = DCCPO_ACK_VECTOR_0;
95         *to++ = len;
96
97         len  = av->dccpav_vec_len;
98         from = av->dccpav_buf + av->dccpav_buf_head;
99
100         /* Check if buf_head wraps */
101         if ((int)av->dccpav_buf_head + len > DCCP_MAX_ACKVEC_LEN) {
102                 const u32 tailsize = DCCP_MAX_ACKVEC_LEN - av->dccpav_buf_head;
103
104                 memcpy(to, from, tailsize);
105                 to   += tailsize;
106                 len  -= tailsize;
107                 from = av->dccpav_buf;
108         }
109
110         memcpy(to, from, len);
111         /*
112          *      From RFC 4340, A.2:
113          *
114          *      For each acknowledgement it sends, the HC-Receiver will add an
115          *      acknowledgement record.  ack_seqno will equal the HC-Receiver
116          *      sequence number it used for the ack packet; ack_ptr will equal
117          *      buf_head; ack_ackno will equal buf_ackno; and ack_nonce will
118          *      equal buf_nonce.
119          */
120         avr->dccpavr_ack_seqno = DCCP_SKB_CB(skb)->dccpd_seq;
121         avr->dccpavr_ack_ptr   = av->dccpav_buf_head;
122         avr->dccpavr_ack_ackno = av->dccpav_buf_ackno;
123         avr->dccpavr_ack_nonce = av->dccpav_buf_nonce;
124         avr->dccpavr_sent_len  = av->dccpav_vec_len;
125
126         dccp_ackvec_insert_avr(av, avr);
127
128         dccp_pr_debug("%s ACK Vector 0, len=%d, ack_seqno=%llu, "
129                       "ack_ackno=%llu\n",
130                       dccp_role(sk), avr->dccpavr_sent_len,
131                       (unsigned long long)avr->dccpavr_ack_seqno,
132                       (unsigned long long)avr->dccpavr_ack_ackno);
133         return 0;
134 }
135
136 struct dccp_ackvec *dccp_ackvec_alloc(const gfp_t priority)
137 {
138         struct dccp_ackvec *av = kmem_cache_alloc(dccp_ackvec_slab, priority);
139
140         if (av != NULL) {
141                 av->dccpav_buf_head     = DCCP_MAX_ACKVEC_LEN - 1;
142                 av->dccpav_buf_ackno    = DCCP_MAX_SEQNO + 1;
143                 av->dccpav_buf_nonce = av->dccpav_buf_nonce = 0;
144                 av->dccpav_ack_ptr      = 0;
145                 av->dccpav_time.tv_sec  = 0;
146                 av->dccpav_time.tv_usec = 0;
147                 av->dccpav_vec_len      = 0;
148                 INIT_LIST_HEAD(&av->dccpav_records);
149         }
150
151         return av;
152 }
153
154 void dccp_ackvec_free(struct dccp_ackvec *av)
155 {
156         if (unlikely(av == NULL))
157                 return;
158
159         if (!list_empty(&av->dccpav_records)) {
160                 struct dccp_ackvec_record *avr, *next;
161
162                 list_for_each_entry_safe(avr, next, &av->dccpav_records,
163                                          dccpavr_node) {
164                         list_del_init(&avr->dccpavr_node);
165                         dccp_ackvec_record_delete(avr);
166                 }
167         }
168
169         kmem_cache_free(dccp_ackvec_slab, av);
170 }
171
172 static inline u8 dccp_ackvec_state(const struct dccp_ackvec *av,
173                                    const u8 index)
174 {
175         return av->dccpav_buf[index] & DCCP_ACKVEC_STATE_MASK;
176 }
177
178 static inline u8 dccp_ackvec_len(const struct dccp_ackvec *av,
179                                  const u8 index)
180 {
181         return av->dccpav_buf[index] & DCCP_ACKVEC_LEN_MASK;
182 }
183
184 /*
185  * If several packets are missing, the HC-Receiver may prefer to enter multiple
186  * bytes with run length 0, rather than a single byte with a larger run length;
187  * this simplifies table updates if one of the missing packets arrives.
188  */
189 static inline int dccp_ackvec_set_buf_head_state(struct dccp_ackvec *av,
190                                                  const unsigned int packets,
191                                                  const unsigned char state)
192 {
193         unsigned int gap;
194         long new_head;
195
196         if (av->dccpav_vec_len + packets > DCCP_MAX_ACKVEC_LEN)
197                 return -ENOBUFS;
198
199         gap      = packets - 1;
200         new_head = av->dccpav_buf_head - packets;
201
202         if (new_head < 0) {
203                 if (gap > 0) {
204                         memset(av->dccpav_buf, DCCP_ACKVEC_STATE_NOT_RECEIVED,
205                                gap + new_head + 1);
206                         gap = -new_head;
207                 }
208                 new_head += DCCP_MAX_ACKVEC_LEN;
209         } 
210
211         av->dccpav_buf_head = new_head;
212
213         if (gap > 0)
214                 memset(av->dccpav_buf + av->dccpav_buf_head + 1,
215                        DCCP_ACKVEC_STATE_NOT_RECEIVED, gap);
216
217         av->dccpav_buf[av->dccpav_buf_head] = state;
218         av->dccpav_vec_len += packets;
219         return 0;
220 }
221
222 /*
223  * Implements the RFC 4340, Appendix A
224  */
225 int dccp_ackvec_add(struct dccp_ackvec *av, const struct sock *sk,
226                     const u64 ackno, const u8 state)
227 {
228         /*
229          * Check at the right places if the buffer is full, if it is, tell the
230          * caller to start dropping packets till the HC-Sender acks our ACK
231          * vectors, when we will free up space in dccpav_buf.
232          *
233          * We may well decide to do buffer compression, etc, but for now lets
234          * just drop.
235          *
236          * From Appendix A.1.1 (`New Packets'):
237          *
238          *      Of course, the circular buffer may overflow, either when the
239          *      HC-Sender is sending data at a very high rate, when the
240          *      HC-Receiver's acknowledgements are not reaching the HC-Sender,
241          *      or when the HC-Sender is forgetting to acknowledge those acks
242          *      (so the HC-Receiver is unable to clean up old state). In this
243          *      case, the HC-Receiver should either compress the buffer (by
244          *      increasing run lengths when possible), transfer its state to
245          *      a larger buffer, or, as a last resort, drop all received
246          *      packets, without processing them whatsoever, until its buffer
247          *      shrinks again.
248          */
249
250         /* See if this is the first ackno being inserted */
251         if (av->dccpav_vec_len == 0) {
252                 av->dccpav_buf[av->dccpav_buf_head] = state;
253                 av->dccpav_vec_len = 1;
254         } else if (after48(ackno, av->dccpav_buf_ackno)) {
255                 const u64 delta = dccp_delta_seqno(av->dccpav_buf_ackno,
256                                                    ackno);
257
258                 /*
259                  * Look if the state of this packet is the same as the
260                  * previous ackno and if so if we can bump the head len.
261                  */
262                 if (delta == 1 &&
263                     dccp_ackvec_state(av, av->dccpav_buf_head) == state &&
264                     (dccp_ackvec_len(av, av->dccpav_buf_head) <
265                      DCCP_ACKVEC_LEN_MASK))
266                         av->dccpav_buf[av->dccpav_buf_head]++;
267                 else if (dccp_ackvec_set_buf_head_state(av, delta, state))
268                         return -ENOBUFS;
269         } else {
270                 /*
271                  * A.1.2.  Old Packets
272                  *
273                  *      When a packet with Sequence Number S <= buf_ackno
274                  *      arrives, the HC-Receiver will scan the table for
275                  *      the byte corresponding to S. (Indexing structures
276                  *      could reduce the complexity of this scan.)
277                  */
278                 u64 delta = dccp_delta_seqno(ackno, av->dccpav_buf_ackno);
279                 u8 index = av->dccpav_buf_head;
280
281                 while (1) {
282                         const u8 len = dccp_ackvec_len(av, index);
283                         const u8 state = dccp_ackvec_state(av, index);
284                         /*
285                          * valid packets not yet in dccpav_buf have a reserved
286                          * entry, with a len equal to 0.
287                          */
288                         if (state == DCCP_ACKVEC_STATE_NOT_RECEIVED &&
289                             len == 0 && delta == 0) { /* Found our
290                                                          reserved seat! */
291                                 dccp_pr_debug("Found %llu reserved seat!\n",
292                                               (unsigned long long)ackno);
293                                 av->dccpav_buf[index] = state;
294                                 goto out;
295                         }
296                         /* len == 0 means one packet */
297                         if (delta < len + 1)
298                                 goto out_duplicate;
299
300                         delta -= len + 1;
301                         if (++index == DCCP_MAX_ACKVEC_LEN)
302                                 index = 0;
303                 }
304         }
305
306         av->dccpav_buf_ackno = ackno;
307         dccp_timestamp(sk, &av->dccpav_time);
308 out:
309         return 0;
310
311 out_duplicate:
312         /* Duplicate packet */
313         dccp_pr_debug("Received a dup or already considered lost "
314                       "packet: %llu\n", (unsigned long long)ackno);
315         return -EILSEQ;
316 }
317
318 #ifdef CONFIG_IP_DCCP_DEBUG
319 void dccp_ackvector_print(const u64 ackno, const unsigned char *vector, int len)
320 {
321         if (!dccp_debug)
322                 return;
323
324         printk("ACK vector len=%d, ackno=%llu |", len,
325                (unsigned long long)ackno);
326
327         while (len--) {
328                 const u8 state = (*vector & DCCP_ACKVEC_STATE_MASK) >> 6;
329                 const u8 rl = *vector & DCCP_ACKVEC_LEN_MASK;
330
331                 printk("%d,%d|", state, rl);
332                 ++vector;
333         }
334
335         printk("\n");
336 }
337
338 void dccp_ackvec_print(const struct dccp_ackvec *av)
339 {
340         dccp_ackvector_print(av->dccpav_buf_ackno,
341                              av->dccpav_buf + av->dccpav_buf_head,
342                              av->dccpav_vec_len);
343 }
344 #endif
345
346 static void dccp_ackvec_throw_record(struct dccp_ackvec *av,
347                                      struct dccp_ackvec_record *avr)
348 {
349         struct dccp_ackvec_record *next;
350
351         /* sort out vector length */
352         if (av->dccpav_buf_head <= avr->dccpavr_ack_ptr)
353                 av->dccpav_vec_len = avr->dccpavr_ack_ptr - av->dccpav_buf_head;
354         else
355                 av->dccpav_vec_len = DCCP_MAX_ACKVEC_LEN - 1
356                                      - av->dccpav_buf_head
357                                      + avr->dccpavr_ack_ptr;
358
359         /* free records */
360         list_for_each_entry_safe_from(avr, next, &av->dccpav_records,
361                                       dccpavr_node) {
362                 list_del_init(&avr->dccpavr_node);
363                 dccp_ackvec_record_delete(avr);
364         }
365 }
366
367 void dccp_ackvec_check_rcv_ackno(struct dccp_ackvec *av, struct sock *sk,
368                                  const u64 ackno)
369 {
370         struct dccp_ackvec_record *avr;
371
372         /*
373          * If we traverse backwards, it should be faster when we have large
374          * windows. We will be receiving ACKs for stuff we sent a while back
375          * -sorbo.
376          */
377         list_for_each_entry_reverse(avr, &av->dccpav_records, dccpavr_node) {
378                 if (ackno == avr->dccpavr_ack_seqno) {
379                         dccp_pr_debug("%s ACK packet 0, len=%d, ack_seqno=%llu, "
380                                       "ack_ackno=%llu, ACKED!\n",
381                                       dccp_role(sk), 1,
382                                       (unsigned long long)avr->dccpavr_ack_seqno,
383                                       (unsigned long long)avr->dccpavr_ack_ackno);
384                         dccp_ackvec_throw_record(av, avr);
385                         break;
386                 } else if (avr->dccpavr_ack_seqno > ackno)
387                         break; /* old news */
388         }
389 }
390
391 static void dccp_ackvec_check_rcv_ackvector(struct dccp_ackvec *av,
392                                             struct sock *sk, u64 ackno,
393                                             const unsigned char len,
394                                             const unsigned char *vector)
395 {
396         unsigned char i;
397         struct dccp_ackvec_record *avr;
398
399         /* Check if we actually sent an ACK vector */
400         if (list_empty(&av->dccpav_records))
401                 return;
402
403         i = len;
404         /*
405          * XXX
406          * I think it might be more efficient to work backwards. See comment on
407          * rcv_ackno. -sorbo.
408          */
409         avr = list_entry(av->dccpav_records.next, struct dccp_ackvec_record,
410                          dccpavr_node);
411         while (i--) {
412                 const u8 rl = *vector & DCCP_ACKVEC_LEN_MASK;
413                 u64 ackno_end_rl;
414
415                 dccp_set_seqno(&ackno_end_rl, ackno - rl);
416
417                 /*
418                  * If our AVR sequence number is greater than the ack, go
419                  * forward in the AVR list until it is not so.
420                  */
421                 list_for_each_entry_from(avr, &av->dccpav_records,
422                                          dccpavr_node) {
423                         if (!after48(avr->dccpavr_ack_seqno, ackno))
424                                 goto found;
425                 }
426                 /* End of the dccpav_records list, not found, exit */
427                 break;
428 found:
429                 if (between48(avr->dccpavr_ack_seqno, ackno_end_rl, ackno)) {
430                         const u8 state = *vector & DCCP_ACKVEC_STATE_MASK;
431                         if (state != DCCP_ACKVEC_STATE_NOT_RECEIVED) {
432                                 dccp_pr_debug("%s ACK vector 0, len=%d, "
433                                               "ack_seqno=%llu, ack_ackno=%llu, "
434                                               "ACKED!\n",
435                                               dccp_role(sk), len,
436                                               (unsigned long long)
437                                               avr->dccpavr_ack_seqno,
438                                               (unsigned long long)
439                                               avr->dccpavr_ack_ackno);
440                                 dccp_ackvec_throw_record(av, avr);
441                                 break;
442                         }
443                         /*
444                          * If it wasn't received, continue scanning... we might
445                          * find another one.
446                          */
447                 }
448
449                 dccp_set_seqno(&ackno, ackno_end_rl - 1);
450                 ++vector;
451         }
452 }
453
454 int dccp_ackvec_parse(struct sock *sk, const struct sk_buff *skb,
455                       const u8 opt, const u8 *value, const u8 len)
456 {
457         if (len > DCCP_MAX_ACKVEC_LEN)
458                 return -1;
459
460         /* dccp_ackvector_print(DCCP_SKB_CB(skb)->dccpd_ack_seq, value, len); */
461         dccp_ackvec_check_rcv_ackvector(dccp_sk(sk)->dccps_hc_rx_ackvec, sk,
462                                         DCCP_SKB_CB(skb)->dccpd_ack_seq,
463                                         len, value);
464         return 0;
465 }
466
467 static char dccp_ackvec_slab_msg[] __initdata =
468         KERN_CRIT "DCCP: Unable to create ack vectors slab caches\n";
469
470 int __init dccp_ackvec_init(void)
471 {
472         dccp_ackvec_slab = kmem_cache_create("dccp_ackvec",
473                                              sizeof(struct dccp_ackvec), 0,
474                                              SLAB_HWCACHE_ALIGN, NULL, NULL);
475         if (dccp_ackvec_slab == NULL)
476                 goto out_err;
477
478         dccp_ackvec_record_slab =
479                         kmem_cache_create("dccp_ackvec_record",
480                                           sizeof(struct dccp_ackvec_record),
481                                           0, SLAB_HWCACHE_ALIGN, NULL, NULL);
482         if (dccp_ackvec_record_slab == NULL)
483                 goto out_destroy_slab;
484
485         return 0;
486
487 out_destroy_slab:
488         kmem_cache_destroy(dccp_ackvec_slab);
489         dccp_ackvec_slab = NULL;
490 out_err:
491         printk(dccp_ackvec_slab_msg);
492         return -ENOBUFS;
493 }
494
495 void dccp_ackvec_exit(void)
496 {
497         if (dccp_ackvec_slab != NULL) {
498                 kmem_cache_destroy(dccp_ackvec_slab);
499                 dccp_ackvec_slab = NULL;
500         }
501         if (dccp_ackvec_record_slab != NULL) {
502                 kmem_cache_destroy(dccp_ackvec_record_slab);
503                 dccp_ackvec_record_slab = NULL;
504         }
505 }