Linux 2.6.31-rc6
[linux-2.6] / net / dccp / ccids / lib / packet_history.c
1 /*
2  *  net/dccp/packet_history.c
3  *
4  *  Copyright (c) 2007   The University of Aberdeen, Scotland, UK
5  *  Copyright (c) 2005-7 The University of Waikato, Hamilton, New Zealand.
6  *
7  *  An implementation of the DCCP protocol
8  *
9  *  This code has been developed by the University of Waikato WAND
10  *  research group. For further information please see http://www.wand.net.nz/
11  *  or e-mail Ian McDonald - ian.mcdonald@jandi.co.nz
12  *
13  *  This code also uses code from Lulea University, rereleased as GPL by its
14  *  authors:
15  *  Copyright (c) 2003 Nils-Erik Mattsson, Joacim Haggmark, Magnus Erixzon
16  *
17  *  Changes to meet Linux coding standards, to make it meet latest ccid3 draft
18  *  and to make it work as a loadable module in the DCCP stack written by
19  *  Arnaldo Carvalho de Melo <acme@conectiva.com.br>.
20  *
21  *  Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@conectiva.com.br>
22  *
23  *  This program is free software; you can redistribute it and/or modify
24  *  it under the terms of the GNU General Public License as published by
25  *  the Free Software Foundation; either version 2 of the License, or
26  *  (at your option) any later version.
27  *
28  *  This program is distributed in the hope that it will be useful,
29  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
30  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
31  *  GNU General Public License for more details.
32  *
33  *  You should have received a copy of the GNU General Public License
34  *  along with this program; if not, write to the Free Software
35  *  Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
36  */
37
38 #include <linux/string.h>
39 #include <linux/slab.h>
40 #include "packet_history.h"
41 #include "../../dccp.h"
42
43 /**
44  *  tfrc_tx_hist_entry  -  Simple singly-linked TX history list
45  *  @next:  next oldest entry (LIFO order)
46  *  @seqno: sequence number of this entry
47  *  @stamp: send time of packet with sequence number @seqno
48  */
49 struct tfrc_tx_hist_entry {
50         struct tfrc_tx_hist_entry *next;
51         u64                       seqno;
52         ktime_t                   stamp;
53 };
54
55 /*
56  * Transmitter History Routines
57  */
58 static struct kmem_cache *tfrc_tx_hist_slab;
59
60 int __init tfrc_tx_packet_history_init(void)
61 {
62         tfrc_tx_hist_slab = kmem_cache_create("tfrc_tx_hist",
63                                               sizeof(struct tfrc_tx_hist_entry),
64                                               0, SLAB_HWCACHE_ALIGN, NULL);
65         return tfrc_tx_hist_slab == NULL ? -ENOBUFS : 0;
66 }
67
68 void tfrc_tx_packet_history_exit(void)
69 {
70         if (tfrc_tx_hist_slab != NULL) {
71                 kmem_cache_destroy(tfrc_tx_hist_slab);
72                 tfrc_tx_hist_slab = NULL;
73         }
74 }
75
76 static struct tfrc_tx_hist_entry *
77         tfrc_tx_hist_find_entry(struct tfrc_tx_hist_entry *head, u64 seqno)
78 {
79         while (head != NULL && head->seqno != seqno)
80                 head = head->next;
81
82         return head;
83 }
84
85 int tfrc_tx_hist_add(struct tfrc_tx_hist_entry **headp, u64 seqno)
86 {
87         struct tfrc_tx_hist_entry *entry = kmem_cache_alloc(tfrc_tx_hist_slab, gfp_any());
88
89         if (entry == NULL)
90                 return -ENOBUFS;
91         entry->seqno = seqno;
92         entry->stamp = ktime_get_real();
93         entry->next  = *headp;
94         *headp       = entry;
95         return 0;
96 }
97
98 void tfrc_tx_hist_purge(struct tfrc_tx_hist_entry **headp)
99 {
100         struct tfrc_tx_hist_entry *head = *headp;
101
102         while (head != NULL) {
103                 struct tfrc_tx_hist_entry *next = head->next;
104
105                 kmem_cache_free(tfrc_tx_hist_slab, head);
106                 head = next;
107         }
108
109         *headp = NULL;
110 }
111
112 u32 tfrc_tx_hist_rtt(struct tfrc_tx_hist_entry *head, const u64 seqno,
113                      const ktime_t now)
114 {
115         u32 rtt = 0;
116         struct tfrc_tx_hist_entry *packet = tfrc_tx_hist_find_entry(head, seqno);
117
118         if (packet != NULL) {
119                 rtt = ktime_us_delta(now, packet->stamp);
120                 /*
121                  * Garbage-collect older (irrelevant) entries:
122                  */
123                 tfrc_tx_hist_purge(&packet->next);
124         }
125
126         return rtt;
127 }
128
129
130 /*
131  *      Receiver History Routines
132  */
133 static struct kmem_cache *tfrc_rx_hist_slab;
134
135 int __init tfrc_rx_packet_history_init(void)
136 {
137         tfrc_rx_hist_slab = kmem_cache_create("tfrc_rxh_cache",
138                                               sizeof(struct tfrc_rx_hist_entry),
139                                               0, SLAB_HWCACHE_ALIGN, NULL);
140         return tfrc_rx_hist_slab == NULL ? -ENOBUFS : 0;
141 }
142
143 void tfrc_rx_packet_history_exit(void)
144 {
145         if (tfrc_rx_hist_slab != NULL) {
146                 kmem_cache_destroy(tfrc_rx_hist_slab);
147                 tfrc_rx_hist_slab = NULL;
148         }
149 }
150
151 static inline void tfrc_rx_hist_entry_from_skb(struct tfrc_rx_hist_entry *entry,
152                                                const struct sk_buff *skb,
153                                                const u64 ndp)
154 {
155         const struct dccp_hdr *dh = dccp_hdr(skb);
156
157         entry->tfrchrx_seqno = DCCP_SKB_CB(skb)->dccpd_seq;
158         entry->tfrchrx_ccval = dh->dccph_ccval;
159         entry->tfrchrx_type  = dh->dccph_type;
160         entry->tfrchrx_ndp   = ndp;
161         entry->tfrchrx_tstamp = ktime_get_real();
162 }
163
164 void tfrc_rx_hist_add_packet(struct tfrc_rx_hist *h,
165                              const struct sk_buff *skb,
166                              const u64 ndp)
167 {
168         struct tfrc_rx_hist_entry *entry = tfrc_rx_hist_last_rcv(h);
169
170         tfrc_rx_hist_entry_from_skb(entry, skb, ndp);
171 }
172
173 /* has the packet contained in skb been seen before? */
174 int tfrc_rx_hist_duplicate(struct tfrc_rx_hist *h, struct sk_buff *skb)
175 {
176         const u64 seq = DCCP_SKB_CB(skb)->dccpd_seq;
177         int i;
178
179         if (dccp_delta_seqno(tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno, seq) <= 0)
180                 return 1;
181
182         for (i = 1; i <= h->loss_count; i++)
183                 if (tfrc_rx_hist_entry(h, i)->tfrchrx_seqno == seq)
184                         return 1;
185
186         return 0;
187 }
188
189 static void tfrc_rx_hist_swap(struct tfrc_rx_hist *h, const u8 a, const u8 b)
190 {
191         const u8 idx_a = tfrc_rx_hist_index(h, a),
192                  idx_b = tfrc_rx_hist_index(h, b);
193         struct tfrc_rx_hist_entry *tmp = h->ring[idx_a];
194
195         h->ring[idx_a] = h->ring[idx_b];
196         h->ring[idx_b] = tmp;
197 }
198
199 /*
200  * Private helper functions for loss detection.
201  *
202  * In the descriptions, `Si' refers to the sequence number of entry number i,
203  * whose NDP count is `Ni' (lower case is used for variables).
204  * Note: All __xxx_loss functions expect that a test against duplicates has been
205  *       performed already: the seqno of the skb must not be less than the seqno
206  *       of loss_prev; and it must not equal that of any valid history entry.
207  */
208 static void __do_track_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u64 n1)
209 {
210         u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
211             s1 = DCCP_SKB_CB(skb)->dccpd_seq;
212
213         if (!dccp_loss_free(s0, s1, n1)) {      /* gap between S0 and S1 */
214                 h->loss_count = 1;
215                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n1);
216         }
217 }
218
219 static void __one_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n2)
220 {
221         u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
222             s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
223             s2 = DCCP_SKB_CB(skb)->dccpd_seq;
224
225         if (likely(dccp_delta_seqno(s1, s2) > 0)) {     /* S1  <  S2 */
226                 h->loss_count = 2;
227                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n2);
228                 return;
229         }
230
231         /* S0  <  S2  <  S1 */
232
233         if (dccp_loss_free(s0, s2, n2)) {
234                 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
235
236                 if (dccp_loss_free(s2, s1, n1)) {
237                         /* hole is filled: S0, S2, and S1 are consecutive */
238                         h->loss_count = 0;
239                         h->loss_start = tfrc_rx_hist_index(h, 1);
240                 } else
241                         /* gap between S2 and S1: just update loss_prev */
242                         tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n2);
243
244         } else {        /* gap between S0 and S2 */
245                 /*
246                  * Reorder history to insert S2 between S0 and S1
247                  */
248                 tfrc_rx_hist_swap(h, 0, 3);
249                 h->loss_start = tfrc_rx_hist_index(h, 3);
250                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n2);
251                 h->loss_count = 2;
252         }
253 }
254
255 /* return 1 if a new loss event has been identified */
256 static int __two_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n3)
257 {
258         u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
259             s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
260             s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
261             s3 = DCCP_SKB_CB(skb)->dccpd_seq;
262
263         if (likely(dccp_delta_seqno(s2, s3) > 0)) {     /* S2  <  S3 */
264                 h->loss_count = 3;
265                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 3), skb, n3);
266                 return 1;
267         }
268
269         /* S3  <  S2 */
270
271         if (dccp_delta_seqno(s1, s3) > 0) {             /* S1  <  S3  <  S2 */
272                 /*
273                  * Reorder history to insert S3 between S1 and S2
274                  */
275                 tfrc_rx_hist_swap(h, 2, 3);
276                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n3);
277                 h->loss_count = 3;
278                 return 1;
279         }
280
281         /* S0  <  S3  <  S1 */
282
283         if (dccp_loss_free(s0, s3, n3)) {
284                 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
285
286                 if (dccp_loss_free(s3, s1, n1)) {
287                         /* hole between S0 and S1 filled by S3 */
288                         u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp;
289
290                         if (dccp_loss_free(s1, s2, n2)) {
291                                 /* entire hole filled by S0, S3, S1, S2 */
292                                 h->loss_start = tfrc_rx_hist_index(h, 2);
293                                 h->loss_count = 0;
294                         } else {
295                                 /* gap remains between S1 and S2 */
296                                 h->loss_start = tfrc_rx_hist_index(h, 1);
297                                 h->loss_count = 1;
298                         }
299
300                 } else /* gap exists between S3 and S1, loss_count stays at 2 */
301                         tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n3);
302
303                 return 0;
304         }
305
306         /*
307          * The remaining case:  S0  <  S3  <  S1  <  S2;  gap between S0 and S3
308          * Reorder history to insert S3 between S0 and S1.
309          */
310         tfrc_rx_hist_swap(h, 0, 3);
311         h->loss_start = tfrc_rx_hist_index(h, 3);
312         tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n3);
313         h->loss_count = 3;
314
315         return 1;
316 }
317
318 /* recycle RX history records to continue loss detection if necessary */
319 static void __three_after_loss(struct tfrc_rx_hist *h)
320 {
321         /*
322          * At this stage we know already that there is a gap between S0 and S1
323          * (since S0 was the highest sequence number received before detecting
324          * the loss). To recycle the loss record, it is thus only necessary to
325          * check for other possible gaps between S1/S2 and between S2/S3.
326          */
327         u64 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
328             s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
329             s3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_seqno;
330         u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp,
331             n3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_ndp;
332
333         if (dccp_loss_free(s1, s2, n2)) {
334
335                 if (dccp_loss_free(s2, s3, n3)) {
336                         /* no gap between S2 and S3: entire hole is filled */
337                         h->loss_start = tfrc_rx_hist_index(h, 3);
338                         h->loss_count = 0;
339                 } else {
340                         /* gap between S2 and S3 */
341                         h->loss_start = tfrc_rx_hist_index(h, 2);
342                         h->loss_count = 1;
343                 }
344
345         } else {        /* gap between S1 and S2 */
346                 h->loss_start = tfrc_rx_hist_index(h, 1);
347                 h->loss_count = 2;
348         }
349 }
350
351 /**
352  *  tfrc_rx_handle_loss  -  Loss detection and further processing
353  *  @h:             The non-empty RX history object
354  *  @lh:            Loss Intervals database to update
355  *  @skb:           Currently received packet
356  *  @ndp:           The NDP count belonging to @skb
357  *  @calc_first_li: Caller-dependent computation of first loss interval in @lh
358  *  @sk:            Used by @calc_first_li (see tfrc_lh_interval_add)
359  *  Chooses action according to pending loss, updates LI database when a new
360  *  loss was detected, and does required post-processing. Returns 1 when caller
361  *  should send feedback, 0 otherwise.
362  *  Since it also takes care of reordering during loss detection and updates the
363  *  records accordingly, the caller should not perform any more RX history
364  *  operations when loss_count is greater than 0 after calling this function.
365  */
366 int tfrc_rx_handle_loss(struct tfrc_rx_hist *h,
367                         struct tfrc_loss_hist *lh,
368                         struct sk_buff *skb, const u64 ndp,
369                         u32 (*calc_first_li)(struct sock *), struct sock *sk)
370 {
371         int is_new_loss = 0;
372
373         if (h->loss_count == 0) {
374                 __do_track_loss(h, skb, ndp);
375         } else if (h->loss_count == 1) {
376                 __one_after_loss(h, skb, ndp);
377         } else if (h->loss_count != 2) {
378                 DCCP_BUG("invalid loss_count %d", h->loss_count);
379         } else if (__two_after_loss(h, skb, ndp)) {
380                 /*
381                  * Update Loss Interval database and recycle RX records
382                  */
383                 is_new_loss = tfrc_lh_interval_add(lh, h, calc_first_li, sk);
384                 __three_after_loss(h);
385         }
386         return is_new_loss;
387 }
388
389 int tfrc_rx_hist_alloc(struct tfrc_rx_hist *h)
390 {
391         int i;
392
393         for (i = 0; i <= TFRC_NDUPACK; i++) {
394                 h->ring[i] = kmem_cache_alloc(tfrc_rx_hist_slab, GFP_ATOMIC);
395                 if (h->ring[i] == NULL)
396                         goto out_free;
397         }
398
399         h->loss_count = h->loss_start = 0;
400         return 0;
401
402 out_free:
403         while (i-- != 0) {
404                 kmem_cache_free(tfrc_rx_hist_slab, h->ring[i]);
405                 h->ring[i] = NULL;
406         }
407         return -ENOBUFS;
408 }
409
410 void tfrc_rx_hist_purge(struct tfrc_rx_hist *h)
411 {
412         int i;
413
414         for (i = 0; i <= TFRC_NDUPACK; ++i)
415                 if (h->ring[i] != NULL) {
416                         kmem_cache_free(tfrc_rx_hist_slab, h->ring[i]);
417                         h->ring[i] = NULL;
418                 }
419 }
420
421 /**
422  * tfrc_rx_hist_rtt_last_s - reference entry to compute RTT samples against
423  */
424 static inline struct tfrc_rx_hist_entry *
425                         tfrc_rx_hist_rtt_last_s(const struct tfrc_rx_hist *h)
426 {
427         return h->ring[0];
428 }
429
430 /**
431  * tfrc_rx_hist_rtt_prev_s: previously suitable (wrt rtt_last_s) RTT-sampling entry
432  */
433 static inline struct tfrc_rx_hist_entry *
434                         tfrc_rx_hist_rtt_prev_s(const struct tfrc_rx_hist *h)
435 {
436         return h->ring[h->rtt_sample_prev];
437 }
438
439 /**
440  * tfrc_rx_hist_sample_rtt  -  Sample RTT from timestamp / CCVal
441  * Based on ideas presented in RFC 4342, 8.1. Returns 0 if it was not able
442  * to compute a sample with given data - calling function should check this.
443  */
444 u32 tfrc_rx_hist_sample_rtt(struct tfrc_rx_hist *h, const struct sk_buff *skb)
445 {
446         u32 sample = 0,
447             delta_v = SUB16(dccp_hdr(skb)->dccph_ccval,
448                             tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
449
450         if (delta_v < 1 || delta_v > 4) {       /* unsuitable CCVal delta */
451                 if (h->rtt_sample_prev == 2) {  /* previous candidate stored */
452                         sample = SUB16(tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_ccval,
453                                        tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
454                         if (sample)
455                                 sample = 4 / sample *
456                                          ktime_us_delta(tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_tstamp,
457                                                         tfrc_rx_hist_rtt_last_s(h)->tfrchrx_tstamp);
458                         else    /*
459                                  * FIXME: This condition is in principle not
460                                  * possible but occurs when CCID is used for
461                                  * two-way data traffic. I have tried to trace
462                                  * it, but the cause does not seem to be here.
463                                  */
464                                 DCCP_BUG("please report to dccp@vger.kernel.org"
465                                          " => prev = %u, last = %u",
466                                          tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_ccval,
467                                          tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
468                 } else if (delta_v < 1) {
469                         h->rtt_sample_prev = 1;
470                         goto keep_ref_for_next_time;
471                 }
472
473         } else if (delta_v == 4) /* optimal match */
474                 sample = ktime_to_us(net_timedelta(tfrc_rx_hist_rtt_last_s(h)->tfrchrx_tstamp));
475         else {                   /* suboptimal match */
476                 h->rtt_sample_prev = 2;
477                 goto keep_ref_for_next_time;
478         }
479
480         if (unlikely(sample > DCCP_SANE_RTT_MAX)) {
481                 DCCP_WARN("RTT sample %u too large, using max\n", sample);
482                 sample = DCCP_SANE_RTT_MAX;
483         }
484
485         h->rtt_sample_prev = 0;        /* use current entry as next reference */
486 keep_ref_for_next_time:
487
488         return sample;
489 }