dccp ccid-3: Update the computation of X_recv
[deliverable/linux.git] / 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 * Transmitter History Routines
45 */
46 static struct kmem_cache *tfrc_tx_hist_slab;
47
48 int __init tfrc_tx_packet_history_init(void)
49 {
50 tfrc_tx_hist_slab = kmem_cache_create("tfrc_tx_hist",
51 sizeof(struct tfrc_tx_hist_entry),
52 0, SLAB_HWCACHE_ALIGN, NULL);
53 return tfrc_tx_hist_slab == NULL ? -ENOBUFS : 0;
54 }
55
56 void tfrc_tx_packet_history_exit(void)
57 {
58 if (tfrc_tx_hist_slab != NULL) {
59 kmem_cache_destroy(tfrc_tx_hist_slab);
60 tfrc_tx_hist_slab = NULL;
61 }
62 }
63
64 int tfrc_tx_hist_add(struct tfrc_tx_hist_entry **headp, u64 seqno)
65 {
66 struct tfrc_tx_hist_entry *entry = kmem_cache_alloc(tfrc_tx_hist_slab, gfp_any());
67
68 if (entry == NULL)
69 return -ENOBUFS;
70 entry->seqno = seqno;
71 entry->stamp = ktime_get_real();
72 entry->next = *headp;
73 *headp = entry;
74 return 0;
75 }
76 EXPORT_SYMBOL_GPL(tfrc_tx_hist_add);
77
78 void tfrc_tx_hist_purge(struct tfrc_tx_hist_entry **headp)
79 {
80 struct tfrc_tx_hist_entry *head = *headp;
81
82 while (head != NULL) {
83 struct tfrc_tx_hist_entry *next = head->next;
84
85 kmem_cache_free(tfrc_tx_hist_slab, head);
86 head = next;
87 }
88
89 *headp = NULL;
90 }
91 EXPORT_SYMBOL_GPL(tfrc_tx_hist_purge);
92
93 /*
94 * Receiver History Routines
95 */
96 static struct kmem_cache *tfrc_rx_hist_slab;
97
98 int __init tfrc_rx_packet_history_init(void)
99 {
100 tfrc_rx_hist_slab = kmem_cache_create("tfrc_rxh_cache",
101 sizeof(struct tfrc_rx_hist_entry),
102 0, SLAB_HWCACHE_ALIGN, NULL);
103 return tfrc_rx_hist_slab == NULL ? -ENOBUFS : 0;
104 }
105
106 void tfrc_rx_packet_history_exit(void)
107 {
108 if (tfrc_rx_hist_slab != NULL) {
109 kmem_cache_destroy(tfrc_rx_hist_slab);
110 tfrc_rx_hist_slab = NULL;
111 }
112 }
113
114 static inline void tfrc_rx_hist_entry_from_skb(struct tfrc_rx_hist_entry *entry,
115 const struct sk_buff *skb,
116 const u64 ndp)
117 {
118 const struct dccp_hdr *dh = dccp_hdr(skb);
119
120 entry->tfrchrx_seqno = DCCP_SKB_CB(skb)->dccpd_seq;
121 entry->tfrchrx_ccval = dh->dccph_ccval;
122 entry->tfrchrx_type = dh->dccph_type;
123 entry->tfrchrx_ndp = ndp;
124 entry->tfrchrx_tstamp = ktime_get_real();
125 }
126
127 void tfrc_rx_hist_add_packet(struct tfrc_rx_hist *h,
128 const struct sk_buff *skb,
129 const u64 ndp)
130 {
131 struct tfrc_rx_hist_entry *entry = tfrc_rx_hist_last_rcv(h);
132
133 tfrc_rx_hist_entry_from_skb(entry, skb, ndp);
134 }
135 EXPORT_SYMBOL_GPL(tfrc_rx_hist_add_packet);
136
137 /* has the packet contained in skb been seen before? */
138 int tfrc_rx_hist_duplicate(struct tfrc_rx_hist *h, struct sk_buff *skb)
139 {
140 const u64 seq = DCCP_SKB_CB(skb)->dccpd_seq;
141 int i;
142
143 if (dccp_delta_seqno(tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno, seq) <= 0)
144 return 1;
145
146 for (i = 1; i <= h->loss_count; i++)
147 if (tfrc_rx_hist_entry(h, i)->tfrchrx_seqno == seq)
148 return 1;
149
150 return 0;
151 }
152 EXPORT_SYMBOL_GPL(tfrc_rx_hist_duplicate);
153
154
155 static void __tfrc_rx_hist_swap(struct tfrc_rx_hist *h, const u8 a, const u8 b)
156 {
157 struct tfrc_rx_hist_entry *tmp = h->ring[a];
158
159 h->ring[a] = h->ring[b];
160 h->ring[b] = tmp;
161 }
162
163 static void tfrc_rx_hist_swap(struct tfrc_rx_hist *h, const u8 a, const u8 b)
164 {
165 __tfrc_rx_hist_swap(h, tfrc_rx_hist_index(h, a),
166 tfrc_rx_hist_index(h, b));
167 }
168
169 /**
170 * tfrc_rx_hist_resume_rtt_sampling - Prepare RX history for RTT sampling
171 * This is called after loss detection has finished, when the history entry
172 * with the index of `loss_count' holds the highest-received sequence number.
173 * RTT sampling requires this information at ring[0] (tfrc_rx_hist_sample_rtt).
174 */
175 static inline void tfrc_rx_hist_resume_rtt_sampling(struct tfrc_rx_hist *h)
176 {
177 __tfrc_rx_hist_swap(h, 0, tfrc_rx_hist_index(h, h->loss_count));
178 h->loss_count = h->loss_start = 0;
179 }
180
181 /*
182 * Private helper functions for loss detection.
183 *
184 * In the descriptions, `Si' refers to the sequence number of entry number i,
185 * whose NDP count is `Ni' (lower case is used for variables).
186 * Note: All __xxx_loss functions expect that a test against duplicates has been
187 * performed already: the seqno of the skb must not be less than the seqno
188 * of loss_prev; and it must not equal that of any valid history entry.
189 */
190 static void __do_track_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u64 n1)
191 {
192 u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
193 s1 = DCCP_SKB_CB(skb)->dccpd_seq;
194
195 if (!dccp_loss_free(s0, s1, n1)) { /* gap between S0 and S1 */
196 h->loss_count = 1;
197 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n1);
198 }
199 }
200
201 static void __one_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n2)
202 {
203 u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
204 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
205 s2 = DCCP_SKB_CB(skb)->dccpd_seq;
206
207 if (likely(dccp_delta_seqno(s1, s2) > 0)) { /* S1 < S2 */
208 h->loss_count = 2;
209 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n2);
210 return;
211 }
212
213 /* S0 < S2 < S1 */
214
215 if (dccp_loss_free(s0, s2, n2)) {
216 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
217
218 if (dccp_loss_free(s2, s1, n1)) {
219 /* hole is filled: S0, S2, and S1 are consecutive */
220 tfrc_rx_hist_resume_rtt_sampling(h);
221 } else
222 /* gap between S2 and S1: just update loss_prev */
223 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n2);
224
225 } else { /* gap between S0 and S2 */
226 /*
227 * Reorder history to insert S2 between S0 and S1
228 */
229 tfrc_rx_hist_swap(h, 0, 3);
230 h->loss_start = tfrc_rx_hist_index(h, 3);
231 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n2);
232 h->loss_count = 2;
233 }
234 }
235
236 /* return 1 if a new loss event has been identified */
237 static int __two_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n3)
238 {
239 u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
240 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
241 s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
242 s3 = DCCP_SKB_CB(skb)->dccpd_seq;
243
244 if (likely(dccp_delta_seqno(s2, s3) > 0)) { /* S2 < S3 */
245 h->loss_count = 3;
246 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 3), skb, n3);
247 return 1;
248 }
249
250 /* S3 < S2 */
251
252 if (dccp_delta_seqno(s1, s3) > 0) { /* S1 < S3 < S2 */
253 /*
254 * Reorder history to insert S3 between S1 and S2
255 */
256 tfrc_rx_hist_swap(h, 2, 3);
257 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n3);
258 h->loss_count = 3;
259 return 1;
260 }
261
262 /* S0 < S3 < S1 */
263
264 if (dccp_loss_free(s0, s3, n3)) {
265 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
266
267 if (dccp_loss_free(s3, s1, n1)) {
268 /* hole between S0 and S1 filled by S3 */
269 u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp;
270
271 if (dccp_loss_free(s1, s2, n2)) {
272 /* entire hole filled by S0, S3, S1, S2 */
273 tfrc_rx_hist_resume_rtt_sampling(h);
274 } else {
275 /* gap remains between S1 and S2 */
276 h->loss_start = tfrc_rx_hist_index(h, 1);
277 h->loss_count = 1;
278 }
279
280 } else /* gap exists between S3 and S1, loss_count stays at 2 */
281 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n3);
282
283 return 0;
284 }
285
286 /*
287 * The remaining case: S0 < S3 < S1 < S2; gap between S0 and S3
288 * Reorder history to insert S3 between S0 and S1.
289 */
290 tfrc_rx_hist_swap(h, 0, 3);
291 h->loss_start = tfrc_rx_hist_index(h, 3);
292 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n3);
293 h->loss_count = 3;
294
295 return 1;
296 }
297
298 /* recycle RX history records to continue loss detection if necessary */
299 static void __three_after_loss(struct tfrc_rx_hist *h)
300 {
301 /*
302 * At this stage we know already that there is a gap between S0 and S1
303 * (since S0 was the highest sequence number received before detecting
304 * the loss). To recycle the loss record, it is thus only necessary to
305 * check for other possible gaps between S1/S2 and between S2/S3.
306 */
307 u64 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
308 s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
309 s3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_seqno;
310 u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp,
311 n3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_ndp;
312
313 if (dccp_loss_free(s1, s2, n2)) {
314
315 if (dccp_loss_free(s2, s3, n3)) {
316 /* no gap between S2 and S3: entire hole is filled */
317 tfrc_rx_hist_resume_rtt_sampling(h);
318 } else {
319 /* gap between S2 and S3 */
320 h->loss_start = tfrc_rx_hist_index(h, 2);
321 h->loss_count = 1;
322 }
323
324 } else { /* gap between S1 and S2 */
325 h->loss_start = tfrc_rx_hist_index(h, 1);
326 h->loss_count = 2;
327 }
328 }
329
330 /**
331 * tfrc_rx_handle_loss - Loss detection and further processing
332 * @h: The non-empty RX history object
333 * @lh: Loss Intervals database to update
334 * @skb: Currently received packet
335 * @ndp: The NDP count belonging to @skb
336 * @calc_first_li: Caller-dependent computation of first loss interval in @lh
337 * @sk: Used by @calc_first_li (see tfrc_lh_interval_add)
338 * Chooses action according to pending loss, updates LI database when a new
339 * loss was detected, and does required post-processing. Returns 1 when caller
340 * should send feedback, 0 otherwise.
341 * Since it also takes care of reordering during loss detection and updates the
342 * records accordingly, the caller should not perform any more RX history
343 * operations when loss_count is greater than 0 after calling this function.
344 */
345 int tfrc_rx_handle_loss(struct tfrc_rx_hist *h,
346 struct tfrc_loss_hist *lh,
347 struct sk_buff *skb, const u64 ndp,
348 u32 (*calc_first_li)(struct sock *), struct sock *sk)
349 {
350 int is_new_loss = 0;
351
352 if (tfrc_rx_hist_duplicate(h, skb))
353 return 0;
354
355 if (h->loss_count == 0) {
356 __do_track_loss(h, skb, ndp);
357 tfrc_rx_hist_sample_rtt(h, skb);
358 } else if (h->loss_count == 1) {
359 __one_after_loss(h, skb, ndp);
360 } else if (h->loss_count != 2) {
361 DCCP_BUG("invalid loss_count %d", h->loss_count);
362 } else if (__two_after_loss(h, skb, ndp)) {
363 /*
364 * Update Loss Interval database and recycle RX records
365 */
366 is_new_loss = tfrc_lh_interval_add(lh, h, calc_first_li, sk);
367 __three_after_loss(h);
368 }
369
370 /*
371 * Update moving-average of `s' and the sum of received payload bytes.
372 */
373 if (dccp_data_packet(skb)) {
374 const u32 payload = skb->len - dccp_hdr(skb)->dccph_doff * 4;
375
376 h->packet_size = tfrc_ewma(h->packet_size, payload, 9);
377 h->bytes_recvd += payload;
378 }
379
380 /* RFC 3448, 6.1: update I_0, whose growth implies p <= p_prev */
381 if (!is_new_loss)
382 tfrc_lh_update_i_mean(lh, skb);
383
384 return is_new_loss;
385 }
386 EXPORT_SYMBOL_GPL(tfrc_rx_handle_loss);
387
388 /* Compute the sending rate X_recv measured between feedback intervals */
389 u32 tfrc_rx_hist_x_recv(struct tfrc_rx_hist *h, const u32 last_x_recv)
390 {
391 u64 bytes = h->bytes_recvd, last_rtt = h->rtt_estimate;
392 s64 delta = ktime_to_us(net_timedelta(h->bytes_start));
393
394 WARN_ON(delta <= 0);
395 /*
396 * Ensure that the sampling interval for X_recv is at least one RTT,
397 * by extending the sampling interval backwards in time, over the last
398 * R_(m-1) seconds, as per rfc3448bis-06, 6.2.
399 * To reduce noise (e.g. when the RTT changes often), this is only
400 * done when delta is smaller than RTT/2.
401 */
402 if (last_x_recv > 0 && delta < last_rtt/2) {
403 tfrc_pr_debug("delta < RTT ==> %ld us < %u us\n",
404 (long)delta, (unsigned)last_rtt);
405
406 delta = (bytes ? delta : 0) + last_rtt;
407 bytes += div_u64((u64)last_x_recv * last_rtt, USEC_PER_SEC);
408 }
409
410 if (unlikely(bytes == 0)) {
411 DCCP_WARN("X_recv == 0, using old value of %u\n", last_x_recv);
412 return last_x_recv;
413 }
414 return scaled_div32(bytes, delta);
415 }
416 EXPORT_SYMBOL_GPL(tfrc_rx_hist_x_recv);
417
418 void tfrc_rx_hist_purge(struct tfrc_rx_hist *h)
419 {
420 int i;
421
422 for (i = 0; i <= TFRC_NDUPACK; ++i)
423 if (h->ring[i] != NULL) {
424 kmem_cache_free(tfrc_rx_hist_slab, h->ring[i]);
425 h->ring[i] = NULL;
426 }
427 }
428 EXPORT_SYMBOL_GPL(tfrc_rx_hist_purge);
429
430 static int tfrc_rx_hist_alloc(struct tfrc_rx_hist *h)
431 {
432 int i;
433
434 memset(h, 0, sizeof(*h));
435
436 for (i = 0; i <= TFRC_NDUPACK; i++) {
437 h->ring[i] = kmem_cache_alloc(tfrc_rx_hist_slab, GFP_ATOMIC);
438 if (h->ring[i] == NULL) {
439 tfrc_rx_hist_purge(h);
440 return -ENOBUFS;
441 }
442 }
443 return 0;
444 }
445
446 int tfrc_rx_hist_init(struct tfrc_rx_hist *h, struct sock *sk)
447 {
448 if (tfrc_rx_hist_alloc(h))
449 return -ENOBUFS;
450 /*
451 * Initialise first entry with GSR to start loss detection as early as
452 * possible. Code using this must not use any other fields. The entry
453 * will be overwritten once the CCID updates its received packets.
454 */
455 tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno = dccp_sk(sk)->dccps_gsr;
456 return 0;
457 }
458 EXPORT_SYMBOL_GPL(tfrc_rx_hist_init);
459
460 /**
461 * tfrc_rx_hist_sample_rtt - Sample RTT from timestamp / CCVal
462 * Based on ideas presented in RFC 4342, 8.1. This function expects that no loss
463 * is pending and uses the following history entries (via rtt_sample_prev):
464 * - h->ring[0] contains the most recent history entry prior to @skb;
465 * - h->ring[1] is an unused `dummy' entry when the current difference is 0;
466 */
467 void tfrc_rx_hist_sample_rtt(struct tfrc_rx_hist *h, const struct sk_buff *skb)
468 {
469 struct tfrc_rx_hist_entry *last = h->ring[0];
470 u32 sample, delta_v;
471
472 /*
473 * When not to sample:
474 * - on non-data packets
475 * (RFC 4342, 8.1: CCVal only fully defined for data packets);
476 * - when no data packets have been received yet
477 * (FIXME: using sampled packet size as indicator here);
478 * - as long as there are gaps in the sequence space (pending loss).
479 */
480 if (!dccp_data_packet(skb) || h->packet_size == 0 ||
481 tfrc_rx_hist_loss_pending(h))
482 return;
483
484 h->rtt_sample_prev = 0; /* reset previous candidate */
485
486 delta_v = SUB16(dccp_hdr(skb)->dccph_ccval, last->tfrchrx_ccval);
487 if (delta_v == 0) { /* less than RTT/4 difference */
488 h->rtt_sample_prev = 1;
489 return;
490 }
491 sample = dccp_sane_rtt(ktime_to_us(net_timedelta(last->tfrchrx_tstamp)));
492
493 if (delta_v <= 4) /* between RTT/4 and RTT */
494 sample *= 4 / delta_v;
495 else if (!(sample < h->rtt_estimate && sample > h->rtt_estimate/2))
496 /*
497 * Optimisation: CCVal difference is greater than 1 RTT, yet the
498 * sample is less than the local RTT estimate; which means that
499 * the RTT estimate is too high.
500 * To avoid noise, it is not done if the sample is below RTT/2.
501 */
502 return;
503
504 /* Use a lower weight than usual to increase responsiveness */
505 h->rtt_estimate = tfrc_ewma(h->rtt_estimate, sample, 5);
506 }
507 EXPORT_SYMBOL_GPL(tfrc_rx_hist_sample_rtt);
This page took 0.040721 seconds and 6 git commands to generate.