2 * net/dccp/packet_history.c
4 * Copyright (c) 2007 The University of Aberdeen, Scotland, UK
5 * Copyright (c) 2005-7 The University of Waikato, Hamilton, New Zealand.
7 * An implementation of the DCCP protocol
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
13 * This code also uses code from Lulea University, rereleased as GPL by its
15 * Copyright (c) 2003 Nils-Erik Mattsson, Joacim Haggmark, Magnus Erixzon
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>.
21 * Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@conectiva.com.br>
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.
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.
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.
38 #include <linux/string.h>
39 #include <linux/slab.h>
40 #include "packet_history.h"
41 #include "../../dccp.h"
44 * Transmitter History Routines
46 static struct kmem_cache *tfrc_tx_hist_slab;
48 int __init tfrc_tx_packet_history_init(void)
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;
56 void tfrc_tx_packet_history_exit(void)
58 if (tfrc_tx_hist_slab != NULL) {
59 kmem_cache_destroy(tfrc_tx_hist_slab);
60 tfrc_tx_hist_slab = NULL;
64 int tfrc_tx_hist_add(struct tfrc_tx_hist_entry **headp, u64 seqno)
66 struct tfrc_tx_hist_entry *entry = kmem_cache_alloc(tfrc_tx_hist_slab, gfp_any());
71 entry->stamp = ktime_get_real();
76 EXPORT_SYMBOL_GPL(tfrc_tx_hist_add);
78 void tfrc_tx_hist_purge(struct tfrc_tx_hist_entry **headp)
80 struct tfrc_tx_hist_entry *head = *headp;
82 while (head != NULL) {
83 struct tfrc_tx_hist_entry *next = head->next;
85 kmem_cache_free(tfrc_tx_hist_slab, head);
91 EXPORT_SYMBOL_GPL(tfrc_tx_hist_purge);
94 * Receiver History Routines
96 static struct kmem_cache *tfrc_rx_hist_slab;
98 int __init tfrc_rx_packet_history_init(void)
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;
106 void tfrc_rx_packet_history_exit(void)
108 if (tfrc_rx_hist_slab != NULL) {
109 kmem_cache_destroy(tfrc_rx_hist_slab);
110 tfrc_rx_hist_slab = NULL;
114 static inline void tfrc_rx_hist_entry_from_skb(struct tfrc_rx_hist_entry *entry,
115 const struct sk_buff *skb,
118 const struct dccp_hdr *dh = dccp_hdr(skb);
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();
127 void tfrc_rx_hist_add_packet(struct tfrc_rx_hist *h,
128 const struct sk_buff *skb,
131 struct tfrc_rx_hist_entry *entry = tfrc_rx_hist_last_rcv(h);
133 tfrc_rx_hist_entry_from_skb(entry, skb, ndp);
135 EXPORT_SYMBOL_GPL(tfrc_rx_hist_add_packet);
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)
140 const u64 seq = DCCP_SKB_CB(skb)->dccpd_seq;
143 if (dccp_delta_seqno(tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno, seq) <= 0)
146 for (i = 1; i <= h->loss_count; i++)
147 if (tfrc_rx_hist_entry(h, i)->tfrchrx_seqno == seq)
152 EXPORT_SYMBOL_GPL(tfrc_rx_hist_duplicate);
155 static void __tfrc_rx_hist_swap(struct tfrc_rx_hist *h, const u8 a, const u8 b)
157 struct tfrc_rx_hist_entry *tmp = h->ring[a];
159 h->ring[a] = h->ring[b];
163 static void tfrc_rx_hist_swap(struct tfrc_rx_hist *h, const u8 a, const u8 b)
165 __tfrc_rx_hist_swap(h, tfrc_rx_hist_index(h, a),
166 tfrc_rx_hist_index(h, b));
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).
175 static inline void tfrc_rx_hist_resume_rtt_sampling(struct tfrc_rx_hist *h)
177 __tfrc_rx_hist_swap(h, 0, tfrc_rx_hist_index(h, h->loss_count));
178 h->loss_count = h->loss_start = 0;
182 * Private helper functions for loss detection.
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.
190 static void __do_track_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u64 n1)
192 u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
193 s1 = DCCP_SKB_CB(skb)->dccpd_seq;
195 if (!dccp_loss_free(s0, s1, n1)) /* gap between S0 and S1 */
199 static void __one_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n2)
201 u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
202 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
203 s2 = DCCP_SKB_CB(skb)->dccpd_seq;
205 if (likely(dccp_delta_seqno(s1, s2) > 0)) { /* S1 < S2 */
207 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n2);
213 if (dccp_loss_free(s0, s2, n2)) {
214 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
216 if (dccp_loss_free(s2, s1, n1)) {
217 /* hole is filled: S0, S2, and S1 are consecutive */
218 tfrc_rx_hist_resume_rtt_sampling(h);
220 /* gap between S2 and S1: just update loss_prev */
221 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n2);
223 } else { /* gap between S0 and S2 */
225 * Reorder history to insert S2 between S0 and S1
227 tfrc_rx_hist_swap(h, 0, 3);
228 h->loss_start = tfrc_rx_hist_index(h, 3);
229 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n2);
234 /* return 1 if a new loss event has been identified */
235 static int __two_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n3)
237 u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
238 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
239 s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
240 s3 = DCCP_SKB_CB(skb)->dccpd_seq;
242 if (likely(dccp_delta_seqno(s2, s3) > 0)) { /* S2 < S3 */
244 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 3), skb, n3);
250 if (dccp_delta_seqno(s1, s3) > 0) { /* S1 < S3 < S2 */
252 * Reorder history to insert S3 between S1 and S2
254 tfrc_rx_hist_swap(h, 2, 3);
255 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n3);
262 if (dccp_loss_free(s0, s3, n3)) {
263 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
265 if (dccp_loss_free(s3, s1, n1)) {
266 /* hole between S0 and S1 filled by S3 */
267 u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp;
269 if (dccp_loss_free(s1, s2, n2)) {
270 /* entire hole filled by S0, S3, S1, S2 */
271 tfrc_rx_hist_resume_rtt_sampling(h);
273 /* gap remains between S1 and S2 */
274 h->loss_start = tfrc_rx_hist_index(h, 1);
278 } else /* gap exists between S3 and S1, loss_count stays at 2 */
279 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n3);
285 * The remaining case: S0 < S3 < S1 < S2; gap between S0 and S3
286 * Reorder history to insert S3 between S0 and S1.
288 tfrc_rx_hist_swap(h, 0, 3);
289 h->loss_start = tfrc_rx_hist_index(h, 3);
290 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n3);
296 /* recycle RX history records to continue loss detection if necessary */
297 static void __three_after_loss(struct tfrc_rx_hist *h)
300 * At this stage we know already that there is a gap between S0 and S1
301 * (since S0 was the highest sequence number received before detecting
302 * the loss). To recycle the loss record, it is thus only necessary to
303 * check for other possible gaps between S1/S2 and between S2/S3.
305 u64 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
306 s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
307 s3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_seqno;
308 u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp,
309 n3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_ndp;
311 if (dccp_loss_free(s1, s2, n2)) {
313 if (dccp_loss_free(s2, s3, n3)) {
314 /* no gap between S2 and S3: entire hole is filled */
315 tfrc_rx_hist_resume_rtt_sampling(h);
317 /* gap between S2 and S3 */
318 h->loss_start = tfrc_rx_hist_index(h, 2);
322 } else { /* gap between S1 and S2 */
323 h->loss_start = tfrc_rx_hist_index(h, 1);
329 * tfrc_rx_congestion_event - Loss detection and further processing
330 * @h: The non-empty RX history object
331 * @lh: Loss Intervals database to update
332 * @skb: Currently received packet
333 * @ndp: The NDP count belonging to @skb
334 * @first_li: Caller-dependent computation of first loss interval in @lh
335 * @sk: Used by @calc_first_li (see tfrc_lh_interval_add)
336 * Chooses action according to pending loss, updates LI database when a new
337 * loss was detected, and does required post-processing. Returns 1 when caller
338 * should send feedback, 0 otherwise.
339 * Since it also takes care of reordering during loss detection and updates the
340 * records accordingly, the caller should not perform any more RX history
341 * operations when loss_count is greater than 0 after calling this function.
343 bool tfrc_rx_congestion_event(struct tfrc_rx_hist *h,
344 struct tfrc_loss_hist *lh,
345 struct sk_buff *skb, const u64 ndp,
346 u32 (*first_li)(struct sock *), struct sock *sk)
348 bool new_event = false;
350 if (tfrc_rx_hist_duplicate(h, skb))
353 if (h->loss_count == 0) {
354 __do_track_loss(h, skb, ndp);
355 tfrc_rx_hist_sample_rtt(h, skb);
356 tfrc_rx_hist_add_packet(h, skb, ndp);
357 } else if (h->loss_count == 1) {
358 __one_after_loss(h, skb, ndp);
359 } else if (h->loss_count != 2) {
360 DCCP_BUG("invalid loss_count %d", h->loss_count);
361 } else if (__two_after_loss(h, skb, ndp)) {
363 * Update Loss Interval database and recycle RX records
365 new_event = tfrc_lh_interval_add(lh, h, first_li, sk);
366 __three_after_loss(h);
370 * Update moving-average of `s' and the sum of received payload bytes.
372 if (dccp_data_packet(skb)) {
373 const u32 payload = skb->len - dccp_hdr(skb)->dccph_doff * 4;
375 h->packet_size = tfrc_ewma(h->packet_size, payload, 9);
376 h->bytes_recvd += payload;
379 /* RFC 3448, 6.1: update I_0, whose growth implies p <= p_prev */
381 tfrc_lh_update_i_mean(lh, skb);
385 EXPORT_SYMBOL_GPL(tfrc_rx_congestion_event);
387 /* Compute the sending rate X_recv measured between feedback intervals */
388 u32 tfrc_rx_hist_x_recv(struct tfrc_rx_hist *h, const u32 last_x_recv)
390 u64 bytes = h->bytes_recvd, last_rtt = h->rtt_estimate;
391 s64 delta = ktime_to_us(net_timedelta(h->bytes_start));
395 * Ensure that the sampling interval for X_recv is at least one RTT,
396 * by extending the sampling interval backwards in time, over the last
397 * R_(m-1) seconds, as per rfc3448bis-06, 6.2.
398 * To reduce noise (e.g. when the RTT changes often), this is only
399 * done when delta is smaller than RTT/2.
401 if (last_x_recv > 0 && delta < last_rtt/2) {
402 tfrc_pr_debug("delta < RTT ==> %ld us < %u us\n",
403 (long)delta, (unsigned)last_rtt);
405 delta = (bytes ? delta : 0) + last_rtt;
406 bytes += div_u64((u64)last_x_recv * last_rtt, USEC_PER_SEC);
409 if (unlikely(bytes == 0)) {
410 DCCP_WARN("X_recv == 0, using old value of %u\n", last_x_recv);
413 return scaled_div32(bytes, delta);
415 EXPORT_SYMBOL_GPL(tfrc_rx_hist_x_recv);
417 void tfrc_rx_hist_purge(struct tfrc_rx_hist *h)
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]);
427 EXPORT_SYMBOL_GPL(tfrc_rx_hist_purge);
429 static int tfrc_rx_hist_alloc(struct tfrc_rx_hist *h)
433 memset(h, 0, sizeof(*h));
435 for (i = 0; i <= TFRC_NDUPACK; i++) {
436 h->ring[i] = kmem_cache_alloc(tfrc_rx_hist_slab, GFP_ATOMIC);
437 if (h->ring[i] == NULL) {
438 tfrc_rx_hist_purge(h);
445 int tfrc_rx_hist_init(struct tfrc_rx_hist *h, struct sock *sk)
447 if (tfrc_rx_hist_alloc(h))
450 * Initialise first entry with GSR to start loss detection as early as
451 * possible. Code using this must not use any other fields. The entry
452 * will be overwritten once the CCID updates its received packets.
454 tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno = dccp_sk(sk)->dccps_gsr;
457 EXPORT_SYMBOL_GPL(tfrc_rx_hist_init);
460 * tfrc_rx_hist_sample_rtt - Sample RTT from timestamp / CCVal
461 * Based on ideas presented in RFC 4342, 8.1. This function expects that no loss
462 * is pending and uses the following history entries (via rtt_sample_prev):
463 * - h->ring[0] contains the most recent history entry prior to @skb;
464 * - h->ring[1] is an unused `dummy' entry when the current difference is 0;
466 void tfrc_rx_hist_sample_rtt(struct tfrc_rx_hist *h, const struct sk_buff *skb)
468 struct tfrc_rx_hist_entry *last = h->ring[0];
472 * When not to sample:
473 * - on non-data packets
474 * (RFC 4342, 8.1: CCVal only fully defined for data packets);
475 * - when no data packets have been received yet
476 * (FIXME: using sampled packet size as indicator here);
477 * - as long as there are gaps in the sequence space (pending loss).
479 if (!dccp_data_packet(skb) || h->packet_size == 0 ||
480 tfrc_rx_hist_loss_pending(h))
483 h->rtt_sample_prev = 0; /* reset previous candidate */
485 delta_v = SUB16(dccp_hdr(skb)->dccph_ccval, last->tfrchrx_ccval);
486 if (delta_v == 0) { /* less than RTT/4 difference */
487 h->rtt_sample_prev = 1;
490 sample = dccp_sane_rtt(ktime_to_us(net_timedelta(last->tfrchrx_tstamp)));
492 if (delta_v <= 4) /* between RTT/4 and RTT */
493 sample *= 4 / delta_v;
494 else if (!(sample < h->rtt_estimate && sample > h->rtt_estimate/2))
496 * Optimisation: CCVal difference is greater than 1 RTT, yet the
497 * sample is less than the local RTT estimate; which means that
498 * the RTT estimate is too high.
499 * To avoid noise, it is not done if the sample is below RTT/2.
503 /* Use a lower weight than usual to increase responsiveness */
504 h->rtt_estimate = tfrc_ewma(h->rtt_estimate, sample, 5);
506 EXPORT_SYMBOL_GPL(tfrc_rx_hist_sample_rtt);