Merge branches 'x86/xen', 'x86/build', 'x86/microcode', 'x86/mm-debug-v2', 'x86/memor...
[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 EXPORT_SYMBOL_GPL(tfrc_tx_hist_add);
98
99 void tfrc_tx_hist_purge(struct tfrc_tx_hist_entry **headp)
100 {
101         struct tfrc_tx_hist_entry *head = *headp;
102
103         while (head != NULL) {
104                 struct tfrc_tx_hist_entry *next = head->next;
105
106                 kmem_cache_free(tfrc_tx_hist_slab, head);
107                 head = next;
108         }
109
110         *headp = NULL;
111 }
112 EXPORT_SYMBOL_GPL(tfrc_tx_hist_purge);
113
114 u32 tfrc_tx_hist_rtt(struct tfrc_tx_hist_entry *head, const u64 seqno,
115                      const ktime_t now)
116 {
117         u32 rtt = 0;
118         struct tfrc_tx_hist_entry *packet = tfrc_tx_hist_find_entry(head, seqno);
119
120         if (packet != NULL) {
121                 rtt = ktime_us_delta(now, packet->stamp);
122                 /*
123                  * Garbage-collect older (irrelevant) entries:
124                  */
125                 tfrc_tx_hist_purge(&packet->next);
126         }
127
128         return rtt;
129 }
130 EXPORT_SYMBOL_GPL(tfrc_tx_hist_rtt);
131
132
133 /*
134  *      Receiver History Routines
135  */
136 static struct kmem_cache *tfrc_rx_hist_slab;
137
138 int __init tfrc_rx_packet_history_init(void)
139 {
140         tfrc_rx_hist_slab = kmem_cache_create("tfrc_rxh_cache",
141                                               sizeof(struct tfrc_rx_hist_entry),
142                                               0, SLAB_HWCACHE_ALIGN, NULL);
143         return tfrc_rx_hist_slab == NULL ? -ENOBUFS : 0;
144 }
145
146 void tfrc_rx_packet_history_exit(void)
147 {
148         if (tfrc_rx_hist_slab != NULL) {
149                 kmem_cache_destroy(tfrc_rx_hist_slab);
150                 tfrc_rx_hist_slab = NULL;
151         }
152 }
153
154 static inline void tfrc_rx_hist_entry_from_skb(struct tfrc_rx_hist_entry *entry,
155                                                const struct sk_buff *skb,
156                                                const u64 ndp)
157 {
158         const struct dccp_hdr *dh = dccp_hdr(skb);
159
160         entry->tfrchrx_seqno = DCCP_SKB_CB(skb)->dccpd_seq;
161         entry->tfrchrx_ccval = dh->dccph_ccval;
162         entry->tfrchrx_type  = dh->dccph_type;
163         entry->tfrchrx_ndp   = ndp;
164         entry->tfrchrx_tstamp = ktime_get_real();
165 }
166
167 void tfrc_rx_hist_add_packet(struct tfrc_rx_hist *h,
168                              const struct sk_buff *skb,
169                              const u64 ndp)
170 {
171         struct tfrc_rx_hist_entry *entry = tfrc_rx_hist_last_rcv(h);
172
173         tfrc_rx_hist_entry_from_skb(entry, skb, ndp);
174 }
175 EXPORT_SYMBOL_GPL(tfrc_rx_hist_add_packet);
176
177 /* has the packet contained in skb been seen before? */
178 int tfrc_rx_hist_duplicate(struct tfrc_rx_hist *h, struct sk_buff *skb)
179 {
180         const u64 seq = DCCP_SKB_CB(skb)->dccpd_seq;
181         int i;
182
183         if (dccp_delta_seqno(tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno, seq) <= 0)
184                 return 1;
185
186         for (i = 1; i <= h->loss_count; i++)
187                 if (tfrc_rx_hist_entry(h, i)->tfrchrx_seqno == seq)
188                         return 1;
189
190         return 0;
191 }
192 EXPORT_SYMBOL_GPL(tfrc_rx_hist_duplicate);
193
194 static void tfrc_rx_hist_swap(struct tfrc_rx_hist *h, const u8 a, const u8 b)
195 {
196         const u8 idx_a = tfrc_rx_hist_index(h, a),
197                  idx_b = tfrc_rx_hist_index(h, b);
198         struct tfrc_rx_hist_entry *tmp = h->ring[idx_a];
199
200         h->ring[idx_a] = h->ring[idx_b];
201         h->ring[idx_b] = tmp;
202 }
203
204 /*
205  * Private helper functions for loss detection.
206  *
207  * In the descriptions, `Si' refers to the sequence number of entry number i,
208  * whose NDP count is `Ni' (lower case is used for variables).
209  * Note: All __xxx_loss functions expect that a test against duplicates has been
210  *       performed already: the seqno of the skb must not be less than the seqno
211  *       of loss_prev; and it must not equal that of any valid history entry.
212  */
213 static void __do_track_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u64 n1)
214 {
215         u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
216             s1 = DCCP_SKB_CB(skb)->dccpd_seq;
217
218         if (!dccp_loss_free(s0, s1, n1)) {      /* gap between S0 and S1 */
219                 h->loss_count = 1;
220                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n1);
221         }
222 }
223
224 static void __one_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n2)
225 {
226         u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
227             s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
228             s2 = DCCP_SKB_CB(skb)->dccpd_seq;
229
230         if (likely(dccp_delta_seqno(s1, s2) > 0)) {     /* S1  <  S2 */
231                 h->loss_count = 2;
232                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n2);
233                 return;
234         }
235
236         /* S0  <  S2  <  S1 */
237
238         if (dccp_loss_free(s0, s2, n2)) {
239                 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
240
241                 if (dccp_loss_free(s2, s1, n1)) {
242                         /* hole is filled: S0, S2, and S1 are consecutive */
243                         h->loss_count = 0;
244                         h->loss_start = tfrc_rx_hist_index(h, 1);
245                 } else
246                         /* gap between S2 and S1: just update loss_prev */
247                         tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n2);
248
249         } else {        /* gap between S0 and S2 */
250                 /*
251                  * Reorder history to insert S2 between S0 and S1
252                  */
253                 tfrc_rx_hist_swap(h, 0, 3);
254                 h->loss_start = tfrc_rx_hist_index(h, 3);
255                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n2);
256                 h->loss_count = 2;
257         }
258 }
259
260 /* return 1 if a new loss event has been identified */
261 static int __two_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n3)
262 {
263         u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
264             s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
265             s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
266             s3 = DCCP_SKB_CB(skb)->dccpd_seq;
267
268         if (likely(dccp_delta_seqno(s2, s3) > 0)) {     /* S2  <  S3 */
269                 h->loss_count = 3;
270                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 3), skb, n3);
271                 return 1;
272         }
273
274         /* S3  <  S2 */
275
276         if (dccp_delta_seqno(s1, s3) > 0) {             /* S1  <  S3  <  S2 */
277                 /*
278                  * Reorder history to insert S3 between S1 and S2
279                  */
280                 tfrc_rx_hist_swap(h, 2, 3);
281                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n3);
282                 h->loss_count = 3;
283                 return 1;
284         }
285
286         /* S0  <  S3  <  S1 */
287
288         if (dccp_loss_free(s0, s3, n3)) {
289                 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
290
291                 if (dccp_loss_free(s3, s1, n1)) {
292                         /* hole between S0 and S1 filled by S3 */
293                         u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp;
294
295                         if (dccp_loss_free(s1, s2, n2)) {
296                                 /* entire hole filled by S0, S3, S1, S2 */
297                                 h->loss_start = tfrc_rx_hist_index(h, 2);
298                                 h->loss_count = 0;
299                         } else {
300                                 /* gap remains between S1 and S2 */
301                                 h->loss_start = tfrc_rx_hist_index(h, 1);
302                                 h->loss_count = 1;
303                         }
304
305                 } else /* gap exists between S3 and S1, loss_count stays at 2 */
306                         tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n3);
307
308                 return 0;
309         }
310
311         /*
312          * The remaining case:  S0  <  S3  <  S1  <  S2;  gap between S0 and S3
313          * Reorder history to insert S3 between S0 and S1.
314          */
315         tfrc_rx_hist_swap(h, 0, 3);
316         h->loss_start = tfrc_rx_hist_index(h, 3);
317         tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n3);
318         h->loss_count = 3;
319
320         return 1;
321 }
322
323 /* recycle RX history records to continue loss detection if necessary */
324 static void __three_after_loss(struct tfrc_rx_hist *h)
325 {
326         /*
327          * At this stage we know already that there is a gap between S0 and S1
328          * (since S0 was the highest sequence number received before detecting
329          * the loss). To recycle the loss record, it is thus only necessary to
330          * check for other possible gaps between S1/S2 and between S2/S3.
331          */
332         u64 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
333             s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
334             s3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_seqno;
335         u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp,
336             n3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_ndp;
337
338         if (dccp_loss_free(s1, s2, n2)) {
339
340                 if (dccp_loss_free(s2, s3, n3)) {
341                         /* no gap between S2 and S3: entire hole is filled */
342                         h->loss_start = tfrc_rx_hist_index(h, 3);
343                         h->loss_count = 0;
344                 } else {
345                         /* gap between S2 and S3 */
346                         h->loss_start = tfrc_rx_hist_index(h, 2);
347                         h->loss_count = 1;
348                 }
349
350         } else {        /* gap between S1 and S2 */
351                 h->loss_start = tfrc_rx_hist_index(h, 1);
352                 h->loss_count = 2;
353         }
354 }
355
356 /**
357  *  tfrc_rx_handle_loss  -  Loss detection and further processing
358  *  @h:             The non-empty RX history object
359  *  @lh:            Loss Intervals database to update
360  *  @skb:           Currently received packet
361  *  @ndp:           The NDP count belonging to @skb
362  *  @calc_first_li: Caller-dependent computation of first loss interval in @lh
363  *  @sk:            Used by @calc_first_li (see tfrc_lh_interval_add)
364  *  Chooses action according to pending loss, updates LI database when a new
365  *  loss was detected, and does required post-processing. Returns 1 when caller
366  *  should send feedback, 0 otherwise.
367  *  Since it also takes care of reordering during loss detection and updates the
368  *  records accordingly, the caller should not perform any more RX history
369  *  operations when loss_count is greater than 0 after calling this function.
370  */
371 int tfrc_rx_handle_loss(struct tfrc_rx_hist *h,
372                         struct tfrc_loss_hist *lh,
373                         struct sk_buff *skb, const u64 ndp,
374                         u32 (*calc_first_li)(struct sock *), struct sock *sk)
375 {
376         int is_new_loss = 0;
377
378         if (h->loss_count == 0) {
379                 __do_track_loss(h, skb, ndp);
380         } else if (h->loss_count == 1) {
381                 __one_after_loss(h, skb, ndp);
382         } else if (h->loss_count != 2) {
383                 DCCP_BUG("invalid loss_count %d", h->loss_count);
384         } else if (__two_after_loss(h, skb, ndp)) {
385                 /*
386                  * Update Loss Interval database and recycle RX records
387                  */
388                 is_new_loss = tfrc_lh_interval_add(lh, h, calc_first_li, sk);
389                 __three_after_loss(h);
390         }
391         return is_new_loss;
392 }
393 EXPORT_SYMBOL_GPL(tfrc_rx_handle_loss);
394
395 int tfrc_rx_hist_alloc(struct tfrc_rx_hist *h)
396 {
397         int i;
398
399         for (i = 0; i <= TFRC_NDUPACK; i++) {
400                 h->ring[i] = kmem_cache_alloc(tfrc_rx_hist_slab, GFP_ATOMIC);
401                 if (h->ring[i] == NULL)
402                         goto out_free;
403         }
404
405         h->loss_count = h->loss_start = 0;
406         return 0;
407
408 out_free:
409         while (i-- != 0) {
410                 kmem_cache_free(tfrc_rx_hist_slab, h->ring[i]);
411                 h->ring[i] = NULL;
412         }
413         return -ENOBUFS;
414 }
415 EXPORT_SYMBOL_GPL(tfrc_rx_hist_alloc);
416
417 void tfrc_rx_hist_purge(struct tfrc_rx_hist *h)
418 {
419         int i;
420
421         for (i = 0; i <= TFRC_NDUPACK; ++i)
422                 if (h->ring[i] != NULL) {
423                         kmem_cache_free(tfrc_rx_hist_slab, h->ring[i]);
424                         h->ring[i] = NULL;
425                 }
426 }
427 EXPORT_SYMBOL_GPL(tfrc_rx_hist_purge);
428
429 /**
430  * tfrc_rx_hist_rtt_last_s - reference entry to compute RTT samples against
431  */
432 static inline struct tfrc_rx_hist_entry *
433                         tfrc_rx_hist_rtt_last_s(const struct tfrc_rx_hist *h)
434 {
435         return h->ring[0];
436 }
437
438 /**
439  * tfrc_rx_hist_rtt_prev_s: previously suitable (wrt rtt_last_s) RTT-sampling entry
440  */
441 static inline struct tfrc_rx_hist_entry *
442                         tfrc_rx_hist_rtt_prev_s(const struct tfrc_rx_hist *h)
443 {
444         return h->ring[h->rtt_sample_prev];
445 }
446
447 /**
448  * tfrc_rx_hist_sample_rtt  -  Sample RTT from timestamp / CCVal
449  * Based on ideas presented in RFC 4342, 8.1. Returns 0 if it was not able
450  * to compute a sample with given data - calling function should check this.
451  */
452 u32 tfrc_rx_hist_sample_rtt(struct tfrc_rx_hist *h, const struct sk_buff *skb)
453 {
454         u32 sample = 0,
455             delta_v = SUB16(dccp_hdr(skb)->dccph_ccval,
456                             tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
457
458         if (delta_v < 1 || delta_v > 4) {       /* unsuitable CCVal delta */
459                 if (h->rtt_sample_prev == 2) {  /* previous candidate stored */
460                         sample = SUB16(tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_ccval,
461                                        tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
462                         if (sample)
463                                 sample = 4 / sample *
464                                          ktime_us_delta(tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_tstamp,
465                                                         tfrc_rx_hist_rtt_last_s(h)->tfrchrx_tstamp);
466                         else    /*
467                                  * FIXME: This condition is in principle not
468                                  * possible but occurs when CCID is used for
469                                  * two-way data traffic. I have tried to trace
470                                  * it, but the cause does not seem to be here.
471                                  */
472                                 DCCP_BUG("please report to dccp@vger.kernel.org"
473                                          " => prev = %u, last = %u",
474                                          tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_ccval,
475                                          tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
476                 } else if (delta_v < 1) {
477                         h->rtt_sample_prev = 1;
478                         goto keep_ref_for_next_time;
479                 }
480
481         } else if (delta_v == 4) /* optimal match */
482                 sample = ktime_to_us(net_timedelta(tfrc_rx_hist_rtt_last_s(h)->tfrchrx_tstamp));
483         else {                   /* suboptimal match */
484                 h->rtt_sample_prev = 2;
485                 goto keep_ref_for_next_time;
486         }
487
488         if (unlikely(sample > DCCP_SANE_RTT_MAX)) {
489                 DCCP_WARN("RTT sample %u too large, using max\n", sample);
490                 sample = DCCP_SANE_RTT_MAX;
491         }
492
493         h->rtt_sample_prev = 0;        /* use current entry as next reference */
494 keep_ref_for_next_time:
495
496         return sample;
497 }
498 EXPORT_SYMBOL_GPL(tfrc_rx_hist_sample_rtt);