1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (c) 2007 The University of Aberdeen, Scotland, UK
4 * Copyright (c) 2005-7 The University of Waikato, Hamilton, New Zealand.
6 * An implementation of the DCCP protocol
8 * This code has been developed by the University of Waikato WAND
9 * research group. For further information please see https://www.wand.net.nz/
10 * or e-mail Ian McDonald - ian.mcdonald@jandi.co.nz
12 * This code also uses code from Lulea University, rereleased as GPL by its
14 * Copyright (c) 2003 Nils-Erik Mattsson, Joacim Haggmark, Magnus Erixzon
16 * Changes to meet Linux coding standards, to make it meet latest ccid3 draft
17 * and to make it work as a loadable module in the DCCP stack written by
18 * Arnaldo Carvalho de Melo <acme@conectiva.com.br>.
20 * Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@conectiva.com.br>
23 #include <linux/string.h>
24 #include <linux/slab.h>
25 #include "packet_history.h"
26 #include "../../dccp.h"
29 * Transmitter History Routines
31 static struct kmem_cache *tfrc_tx_hist_slab;
33 int __init tfrc_tx_packet_history_init(void)
35 tfrc_tx_hist_slab = kmem_cache_create("tfrc_tx_hist",
36 sizeof(struct tfrc_tx_hist_entry),
37 0, SLAB_HWCACHE_ALIGN, NULL);
38 return tfrc_tx_hist_slab == NULL ? -ENOBUFS : 0;
41 void tfrc_tx_packet_history_exit(void)
43 if (tfrc_tx_hist_slab != NULL) {
44 kmem_cache_destroy(tfrc_tx_hist_slab);
45 tfrc_tx_hist_slab = NULL;
49 int tfrc_tx_hist_add(struct tfrc_tx_hist_entry **headp, u64 seqno)
51 struct tfrc_tx_hist_entry *entry = kmem_cache_alloc(tfrc_tx_hist_slab, gfp_any());
56 entry->stamp = ktime_get_real();
62 void tfrc_tx_hist_purge(struct tfrc_tx_hist_entry **headp)
64 struct tfrc_tx_hist_entry *head = *headp;
66 while (head != NULL) {
67 struct tfrc_tx_hist_entry *next = head->next;
69 kmem_cache_free(tfrc_tx_hist_slab, head);
77 * Receiver History Routines
79 static struct kmem_cache *tfrc_rx_hist_slab;
81 int __init tfrc_rx_packet_history_init(void)
83 tfrc_rx_hist_slab = kmem_cache_create("tfrc_rxh_cache",
84 sizeof(struct tfrc_rx_hist_entry),
85 0, SLAB_HWCACHE_ALIGN, NULL);
86 return tfrc_rx_hist_slab == NULL ? -ENOBUFS : 0;
89 void tfrc_rx_packet_history_exit(void)
91 if (tfrc_rx_hist_slab != NULL) {
92 kmem_cache_destroy(tfrc_rx_hist_slab);
93 tfrc_rx_hist_slab = NULL;
97 static inline void tfrc_rx_hist_entry_from_skb(struct tfrc_rx_hist_entry *entry,
98 const struct sk_buff *skb,
101 const struct dccp_hdr *dh = dccp_hdr(skb);
103 entry->tfrchrx_seqno = DCCP_SKB_CB(skb)->dccpd_seq;
104 entry->tfrchrx_ccval = dh->dccph_ccval;
105 entry->tfrchrx_type = dh->dccph_type;
106 entry->tfrchrx_ndp = ndp;
107 entry->tfrchrx_tstamp = ktime_get_real();
110 void tfrc_rx_hist_add_packet(struct tfrc_rx_hist *h,
111 const struct sk_buff *skb,
114 struct tfrc_rx_hist_entry *entry = tfrc_rx_hist_last_rcv(h);
116 tfrc_rx_hist_entry_from_skb(entry, skb, ndp);
119 /* has the packet contained in skb been seen before? */
120 int tfrc_rx_hist_duplicate(struct tfrc_rx_hist *h, struct sk_buff *skb)
122 const u64 seq = DCCP_SKB_CB(skb)->dccpd_seq;
125 if (dccp_delta_seqno(tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno, seq) <= 0)
128 for (i = 1; i <= h->loss_count; i++)
129 if (tfrc_rx_hist_entry(h, i)->tfrchrx_seqno == seq)
135 static void tfrc_rx_hist_swap(struct tfrc_rx_hist *h, const u8 a, const u8 b)
137 const u8 idx_a = tfrc_rx_hist_index(h, a),
138 idx_b = tfrc_rx_hist_index(h, b);
140 swap(h->ring[idx_a], h->ring[idx_b]);
144 * Private helper functions for loss detection.
146 * In the descriptions, `Si' refers to the sequence number of entry number i,
147 * whose NDP count is `Ni' (lower case is used for variables).
148 * Note: All __xxx_loss functions expect that a test against duplicates has been
149 * performed already: the seqno of the skb must not be less than the seqno
150 * of loss_prev; and it must not equal that of any valid history entry.
152 static void __do_track_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u64 n1)
154 u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
155 s1 = DCCP_SKB_CB(skb)->dccpd_seq;
157 if (!dccp_loss_free(s0, s1, n1)) { /* gap between S0 and S1 */
159 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n1);
163 static void __one_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n2)
165 u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
166 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
167 s2 = DCCP_SKB_CB(skb)->dccpd_seq;
169 if (likely(dccp_delta_seqno(s1, s2) > 0)) { /* S1 < S2 */
171 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n2);
177 if (dccp_loss_free(s0, s2, n2)) {
178 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
180 if (dccp_loss_free(s2, s1, n1)) {
181 /* hole is filled: S0, S2, and S1 are consecutive */
183 h->loss_start = tfrc_rx_hist_index(h, 1);
185 /* gap between S2 and S1: just update loss_prev */
186 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n2);
188 } else { /* gap between S0 and S2 */
190 * Reorder history to insert S2 between S0 and S1
192 tfrc_rx_hist_swap(h, 0, 3);
193 h->loss_start = tfrc_rx_hist_index(h, 3);
194 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n2);
199 /* return 1 if a new loss event has been identified */
200 static int __two_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n3)
202 u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
203 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
204 s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
205 s3 = DCCP_SKB_CB(skb)->dccpd_seq;
207 if (likely(dccp_delta_seqno(s2, s3) > 0)) { /* S2 < S3 */
209 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 3), skb, n3);
215 if (dccp_delta_seqno(s1, s3) > 0) { /* S1 < S3 < S2 */
217 * Reorder history to insert S3 between S1 and S2
219 tfrc_rx_hist_swap(h, 2, 3);
220 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n3);
227 if (dccp_loss_free(s0, s3, n3)) {
228 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
230 if (dccp_loss_free(s3, s1, n1)) {
231 /* hole between S0 and S1 filled by S3 */
232 u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp;
234 if (dccp_loss_free(s1, s2, n2)) {
235 /* entire hole filled by S0, S3, S1, S2 */
236 h->loss_start = tfrc_rx_hist_index(h, 2);
239 /* gap remains between S1 and S2 */
240 h->loss_start = tfrc_rx_hist_index(h, 1);
244 } else /* gap exists between S3 and S1, loss_count stays at 2 */
245 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n3);
251 * The remaining case: S0 < S3 < S1 < S2; gap between S0 and S3
252 * Reorder history to insert S3 between S0 and S1.
254 tfrc_rx_hist_swap(h, 0, 3);
255 h->loss_start = tfrc_rx_hist_index(h, 3);
256 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n3);
262 /* recycle RX history records to continue loss detection if necessary */
263 static void __three_after_loss(struct tfrc_rx_hist *h)
266 * At this stage we know already that there is a gap between S0 and S1
267 * (since S0 was the highest sequence number received before detecting
268 * the loss). To recycle the loss record, it is thus only necessary to
269 * check for other possible gaps between S1/S2 and between S2/S3.
271 u64 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
272 s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
273 s3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_seqno;
274 u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp,
275 n3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_ndp;
277 if (dccp_loss_free(s1, s2, n2)) {
279 if (dccp_loss_free(s2, s3, n3)) {
280 /* no gap between S2 and S3: entire hole is filled */
281 h->loss_start = tfrc_rx_hist_index(h, 3);
284 /* gap between S2 and S3 */
285 h->loss_start = tfrc_rx_hist_index(h, 2);
289 } else { /* gap between S1 and S2 */
290 h->loss_start = tfrc_rx_hist_index(h, 1);
296 * tfrc_rx_handle_loss - Loss detection and further processing
297 * @h: The non-empty RX history object
298 * @lh: Loss Intervals database to update
299 * @skb: Currently received packet
300 * @ndp: The NDP count belonging to @skb
301 * @calc_first_li: Caller-dependent computation of first loss interval in @lh
302 * @sk: Used by @calc_first_li (see tfrc_lh_interval_add)
304 * Chooses action according to pending loss, updates LI database when a new
305 * loss was detected, and does required post-processing. Returns 1 when caller
306 * should send feedback, 0 otherwise.
307 * Since it also takes care of reordering during loss detection and updates the
308 * records accordingly, the caller should not perform any more RX history
309 * operations when loss_count is greater than 0 after calling this function.
311 int tfrc_rx_handle_loss(struct tfrc_rx_hist *h,
312 struct tfrc_loss_hist *lh,
313 struct sk_buff *skb, const u64 ndp,
314 u32 (*calc_first_li)(struct sock *), struct sock *sk)
318 if (h->loss_count == 0) {
319 __do_track_loss(h, skb, ndp);
320 } else if (h->loss_count == 1) {
321 __one_after_loss(h, skb, ndp);
322 } else if (h->loss_count != 2) {
323 DCCP_BUG("invalid loss_count %d", h->loss_count);
324 } else if (__two_after_loss(h, skb, ndp)) {
326 * Update Loss Interval database and recycle RX records
328 is_new_loss = tfrc_lh_interval_add(lh, h, calc_first_li, sk);
329 __three_after_loss(h);
334 int tfrc_rx_hist_alloc(struct tfrc_rx_hist *h)
338 for (i = 0; i <= TFRC_NDUPACK; i++) {
339 h->ring[i] = kmem_cache_alloc(tfrc_rx_hist_slab, GFP_ATOMIC);
340 if (h->ring[i] == NULL)
344 h->loss_count = h->loss_start = 0;
349 kmem_cache_free(tfrc_rx_hist_slab, h->ring[i]);
355 void tfrc_rx_hist_purge(struct tfrc_rx_hist *h)
359 for (i = 0; i <= TFRC_NDUPACK; ++i)
360 if (h->ring[i] != NULL) {
361 kmem_cache_free(tfrc_rx_hist_slab, h->ring[i]);
367 * tfrc_rx_hist_rtt_last_s - reference entry to compute RTT samples against
368 * @h: The non-empty RX history object
370 static inline struct tfrc_rx_hist_entry *
371 tfrc_rx_hist_rtt_last_s(const struct tfrc_rx_hist *h)
377 * tfrc_rx_hist_rtt_prev_s - previously suitable (wrt rtt_last_s) RTT-sampling entry
378 * @h: The non-empty RX history object
380 static inline struct tfrc_rx_hist_entry *
381 tfrc_rx_hist_rtt_prev_s(const struct tfrc_rx_hist *h)
383 return h->ring[h->rtt_sample_prev];
387 * tfrc_rx_hist_sample_rtt - Sample RTT from timestamp / CCVal
388 * Based on ideas presented in RFC 4342, 8.1. Returns 0 if it was not able
389 * to compute a sample with given data - calling function should check this.
391 u32 tfrc_rx_hist_sample_rtt(struct tfrc_rx_hist *h, const struct sk_buff *skb)
394 delta_v = SUB16(dccp_hdr(skb)->dccph_ccval,
395 tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
397 if (delta_v < 1 || delta_v > 4) { /* unsuitable CCVal delta */
398 if (h->rtt_sample_prev == 2) { /* previous candidate stored */
399 sample = SUB16(tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_ccval,
400 tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
402 sample = 4 / sample *
403 ktime_us_delta(tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_tstamp,
404 tfrc_rx_hist_rtt_last_s(h)->tfrchrx_tstamp);
406 * FIXME: This condition is in principle not
407 * possible but occurs when CCID is used for
408 * two-way data traffic. I have tried to trace
409 * it, but the cause does not seem to be here.
411 DCCP_BUG("please report to dccp@vger.kernel.org"
412 " => prev = %u, last = %u",
413 tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_ccval,
414 tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
415 } else if (delta_v < 1) {
416 h->rtt_sample_prev = 1;
417 goto keep_ref_for_next_time;
420 } else if (delta_v == 4) /* optimal match */
421 sample = ktime_to_us(net_timedelta(tfrc_rx_hist_rtt_last_s(h)->tfrchrx_tstamp));
422 else { /* suboptimal match */
423 h->rtt_sample_prev = 2;
424 goto keep_ref_for_next_time;
427 if (unlikely(sample > DCCP_SANE_RTT_MAX)) {
428 DCCP_WARN("RTT sample %u too large, using max\n", sample);
429 sample = DCCP_SANE_RTT_MAX;
432 h->rtt_sample_prev = 0; /* use current entry as next reference */
433 keep_ref_for_next_time: