1/*
2 * CAIA Delay-Gradient (CDG) congestion control
3 *
4 * This implementation is based on the paper:
5 * D.A. Hayes and G. Armitage. "Revisiting TCP congestion control using
6 * delay gradients." In IFIP Networking, pages 328-341. Springer, 2011.
7 *
8 * Scavenger traffic (Less-than-Best-Effort) should disable coexistence
9 * heuristics using parameters use_shadow=0 and use_ineff=0.
10 *
11 * Parameters window, backoff_beta, and backoff_factor are crucial for
12 * throughput and delay. Future work is needed to determine better defaults,
13 * and to provide guidelines for use in different environments/contexts.
14 *
15 * Except for window, knobs are configured via /sys/module/tcp_cdg/parameters/.
16 * Parameter window is only configurable when loading tcp_cdg as a module.
17 *
18 * Notable differences from paper/FreeBSD:
19 * o Using Hybrid Slow start and Proportional Rate Reduction.
20 * o Add toggle for shadow window mechanism. Suggested by David Hayes.
21 * o Add toggle for non-congestion loss tolerance.
22 * o Scaling parameter G is changed to a backoff factor;
23 * conversion is given by: backoff_factor = 1000/(G * window).
24 * o Limit shadow window to 2 * cwnd, or to cwnd when application limited.
25 * o More accurate e^-x.
26 */
27#include <linux/kernel.h>
28#include <linux/random.h>
29#include <linux/module.h>
30#include <linux/sched/clock.h>
31
32#include <net/tcp.h>
33
34#define HYSTART_ACK_TRAIN 1
35#define HYSTART_DELAY 2
36
37static int window __read_mostly = 8;
38static unsigned int backoff_beta __read_mostly = 0.7071 * 1024; /* sqrt 0.5 */
39static unsigned int backoff_factor __read_mostly = 42;
40static unsigned int hystart_detect __read_mostly = 3;
41static unsigned int use_ineff __read_mostly = 5;
42static bool use_shadow __read_mostly = true;
43static bool use_tolerance __read_mostly;
44
45module_param(window, int, 0444);
46MODULE_PARM_DESC(window, "gradient window size (power of two <= 256)");
47module_param(backoff_beta, uint, 0644);
48MODULE_PARM_DESC(backoff_beta, "backoff beta (0-1024)");
49module_param(backoff_factor, uint, 0644);
50MODULE_PARM_DESC(backoff_factor, "backoff probability scale factor");
51module_param(hystart_detect, uint, 0644);
52MODULE_PARM_DESC(hystart_detect, "use Hybrid Slow start "
53 "(0: disabled, 1: ACK train, 2: delay threshold, 3: both)");
54module_param(use_ineff, uint, 0644);
55MODULE_PARM_DESC(use_ineff, "use ineffectual backoff detection (threshold)");
56module_param(use_shadow, bool, 0644);
57MODULE_PARM_DESC(use_shadow, "use shadow window heuristic");
58module_param(use_tolerance, bool, 0644);
59MODULE_PARM_DESC(use_tolerance, "use loss tolerance heuristic");
60
61struct cdg_minmax {
62 union {
63 struct {
64 s32 min;
65 s32 max;
66 };
67 u64 v64;
68 };
69};
70
71enum cdg_state {
72 CDG_UNKNOWN = 0,
73 CDG_NONFULL = 1,
74 CDG_FULL = 2,
75 CDG_BACKOFF = 3,
76};
77
78struct cdg {
79 struct cdg_minmax rtt;
80 struct cdg_minmax rtt_prev;
81 struct cdg_minmax *gradients;
82 struct cdg_minmax gsum;
83 bool gfilled;
84 u8 tail;
85 u8 state;
86 u8 delack;
87 u32 rtt_seq;
88 u32 shadow_wnd;
89 u16 backoff_cnt;
90 u16 sample_cnt;
91 s32 delay_min;
92 u32 last_ack;
93 u32 round_start;
94};
95
96/**
97 * nexp_u32 - negative base-e exponential
98 * @ux: x in units of micro
99 *
100 * Returns exp(ux * -1e-6) * U32_MAX.
101 */
102static u32 __pure nexp_u32(u32 ux)
103{
104 static const u16 v[] = {
105 /* exp(-x)*65536-1 for x = 0, 0.000256, 0.000512, ... */
106 65535,
107 65518, 65501, 65468, 65401, 65267, 65001, 64470, 63422,
108 61378, 57484, 50423, 38795, 22965, 8047, 987, 14,
109 };
110 u32 msb = ux >> 8;
111 u32 res;
112 int i;
113
114 /* Cut off when ux >= 2^24 (actual result is <= 222/U32_MAX). */
115 if (msb > U16_MAX)
116 return 0;
117
118 /* Scale first eight bits linearly: */
119 res = U32_MAX - (ux & 0xff) * (U32_MAX / 1000000);
120
121 /* Obtain e^(x + y + ...) by computing e^x * e^y * ...: */
122 for (i = 1; msb; i++, msb >>= 1) {
123 u32 y = v[i & -(msb & 1)] + U32_C(1);
124
125 res = ((u64)res * y) >> 16;
126 }
127
128 return res;
129}
130
131/* Based on the HyStart algorithm (by Ha et al.) that is implemented in
132 * tcp_cubic. Differences/experimental changes:
133 * o Using Hayes' delayed ACK filter.
134 * o Using a usec clock for the ACK train.
135 * o Reset ACK train when application limited.
136 * o Invoked at any cwnd (i.e. also when cwnd < 16).
137 * o Invoked only when cwnd < ssthresh (i.e. not when cwnd == ssthresh).
138 */
139static void tcp_cdg_hystart_update(struct sock *sk)
140{
141 struct cdg *ca = inet_csk_ca(sk);
142 struct tcp_sock *tp = tcp_sk(sk);
143
144 ca->delay_min = min_not_zero(ca->delay_min, ca->rtt.min);
145 if (ca->delay_min == 0)
146 return;
147
148 if (hystart_detect & HYSTART_ACK_TRAIN) {
149 u32 now_us = tp->tcp_mstamp;
150
151 if (ca->last_ack == 0 || !tcp_is_cwnd_limited(sk)) {
152 ca->last_ack = now_us;
153 ca->round_start = now_us;
154 } else if (before(now_us, ca->last_ack + 3000)) {
155 u32 base_owd = max(ca->delay_min / 2U, 125U);
156
157 ca->last_ack = now_us;
158 if (after(now_us, ca->round_start + base_owd)) {
159 NET_INC_STATS(sock_net(sk),
160 LINUX_MIB_TCPHYSTARTTRAINDETECT);
161 NET_ADD_STATS(sock_net(sk),
162 LINUX_MIB_TCPHYSTARTTRAINCWND,
163 tp->snd_cwnd);
164 tp->snd_ssthresh = tp->snd_cwnd;
165 return;
166 }
167 }
168 }
169
170 if (hystart_detect & HYSTART_DELAY) {
171 if (ca->sample_cnt < 8) {
172 ca->sample_cnt++;
173 } else {
174 s32 thresh = max(ca->delay_min + ca->delay_min / 8U,
175 125U);
176
177 if (ca->rtt.min > thresh) {
178 NET_INC_STATS(sock_net(sk),
179 LINUX_MIB_TCPHYSTARTDELAYDETECT);
180 NET_ADD_STATS(sock_net(sk),
181 LINUX_MIB_TCPHYSTARTDELAYCWND,
182 tp->snd_cwnd);
183 tp->snd_ssthresh = tp->snd_cwnd;
184 }
185 }
186 }
187}
188
189static s32 tcp_cdg_grad(struct cdg *ca)
190{
191 s32 gmin = ca->rtt.min - ca->rtt_prev.min;
192 s32 gmax = ca->rtt.max - ca->rtt_prev.max;
193 s32 grad;
194
195 if (ca->gradients) {
196 ca->gsum.min += gmin - ca->gradients[ca->tail].min;
197 ca->gsum.max += gmax - ca->gradients[ca->tail].max;
198 ca->gradients[ca->tail].min = gmin;
199 ca->gradients[ca->tail].max = gmax;
200 ca->tail = (ca->tail + 1) & (window - 1);
201 gmin = ca->gsum.min;
202 gmax = ca->gsum.max;
203 }
204
205 /* We keep sums to ignore gradients during cwnd reductions;
206 * the paper's smoothed gradients otherwise simplify to:
207 * (rtt_latest - rtt_oldest) / window.
208 *
209 * We also drop division by window here.
210 */
211 grad = gmin > 0 ? gmin : gmax;
212
213 /* Extrapolate missing values in gradient window: */
214 if (!ca->gfilled) {
215 if (!ca->gradients && window > 1)
216 grad *= window; /* Memory allocation failed. */
217 else if (ca->tail == 0)
218 ca->gfilled = true;
219 else
220 grad = (grad * window) / (int)ca->tail;
221 }
222
223 /* Backoff was effectual: */
224 if (gmin <= -32 || gmax <= -32)
225 ca->backoff_cnt = 0;
226
227 if (use_tolerance) {
228 /* Reduce small variations to zero: */
229 gmin = DIV_ROUND_CLOSEST(gmin, 64);
230 gmax = DIV_ROUND_CLOSEST(gmax, 64);
231
232 if (gmin > 0 && gmax <= 0)
233 ca->state = CDG_FULL;
234 else if ((gmin > 0 && gmax > 0) || gmax < 0)
235 ca->state = CDG_NONFULL;
236 }
237 return grad;
238}
239
240static bool tcp_cdg_backoff(struct sock *sk, u32 grad)
241{
242 struct cdg *ca = inet_csk_ca(sk);
243 struct tcp_sock *tp = tcp_sk(sk);
244
245 if (prandom_u32() <= nexp_u32(grad * backoff_factor))
246 return false;
247
248 if (use_ineff) {
249 ca->backoff_cnt++;
250 if (ca->backoff_cnt > use_ineff)
251 return false;
252 }
253
254 ca->shadow_wnd = max(ca->shadow_wnd, tp->snd_cwnd);
255 ca->state = CDG_BACKOFF;
256 tcp_enter_cwr(sk);
257 return true;
258}
259
260/* Not called in CWR or Recovery state. */
261static void tcp_cdg_cong_avoid(struct sock *sk, u32 ack, u32 acked)
262{
263 struct cdg *ca = inet_csk_ca(sk);
264 struct tcp_sock *tp = tcp_sk(sk);
265 u32 prior_snd_cwnd;
266 u32 incr;
267
268 if (tcp_in_slow_start(tp) && hystart_detect)
269 tcp_cdg_hystart_update(sk);
270
271 if (after(ack, ca->rtt_seq) && ca->rtt.v64) {
272 s32 grad = 0;
273
274 if (ca->rtt_prev.v64)
275 grad = tcp_cdg_grad(ca);
276 ca->rtt_seq = tp->snd_nxt;
277 ca->rtt_prev = ca->rtt;
278 ca->rtt.v64 = 0;
279 ca->last_ack = 0;
280 ca->sample_cnt = 0;
281
282 if (grad > 0 && tcp_cdg_backoff(sk, grad))
283 return;
284 }
285
286 if (!tcp_is_cwnd_limited(sk)) {
287 ca->shadow_wnd = min(ca->shadow_wnd, tp->snd_cwnd);
288 return;
289 }
290
291 prior_snd_cwnd = tp->snd_cwnd;
292 tcp_reno_cong_avoid(sk, ack, acked);
293
294 incr = tp->snd_cwnd - prior_snd_cwnd;
295 ca->shadow_wnd = max(ca->shadow_wnd, ca->shadow_wnd + incr);
296}
297
298static void tcp_cdg_acked(struct sock *sk, const struct ack_sample *sample)
299{
300 struct cdg *ca = inet_csk_ca(sk);
301 struct tcp_sock *tp = tcp_sk(sk);
302
303 if (sample->rtt_us <= 0)
304 return;
305
306 /* A heuristic for filtering delayed ACKs, adapted from:
307 * D.A. Hayes. "Timing enhancements to the FreeBSD kernel to support
308 * delay and rate based TCP mechanisms." TR 100219A. CAIA, 2010.
309 */
310 if (tp->sacked_out == 0) {
311 if (sample->pkts_acked == 1 && ca->delack) {
312 /* A delayed ACK is only used for the minimum if it is
313 * provenly lower than an existing non-zero minimum.
314 */
315 ca->rtt.min = min(ca->rtt.min, sample->rtt_us);
316 ca->delack--;
317 return;
318 } else if (sample->pkts_acked > 1 && ca->delack < 5) {
319 ca->delack++;
320 }
321 }
322
323 ca->rtt.min = min_not_zero(ca->rtt.min, sample->rtt_us);
324 ca->rtt.max = max(ca->rtt.max, sample->rtt_us);
325}
326
327static u32 tcp_cdg_ssthresh(struct sock *sk)
328{
329 struct cdg *ca = inet_csk_ca(sk);
330 struct tcp_sock *tp = tcp_sk(sk);
331
332 if (ca->state == CDG_BACKOFF)
333 return max(2U, (tp->snd_cwnd * min(1024U, backoff_beta)) >> 10);
334
335 if (ca->state == CDG_NONFULL && use_tolerance)
336 return tp->snd_cwnd;
337
338 ca->shadow_wnd = min(ca->shadow_wnd >> 1, tp->snd_cwnd);
339 if (use_shadow)
340 return max3(2U, ca->shadow_wnd, tp->snd_cwnd >> 1);
341 return max(2U, tp->snd_cwnd >> 1);
342}
343
344static void tcp_cdg_cwnd_event(struct sock *sk, const enum tcp_ca_event ev)
345{
346 struct cdg *ca = inet_csk_ca(sk);
347 struct tcp_sock *tp = tcp_sk(sk);
348 struct cdg_minmax *gradients;
349
350 switch (ev) {
351 case CA_EVENT_CWND_RESTART:
352 gradients = ca->gradients;
353 if (gradients)
354 memset(gradients, 0, window * sizeof(gradients[0]));
355 memset(ca, 0, sizeof(*ca));
356
357 ca->gradients = gradients;
358 ca->rtt_seq = tp->snd_nxt;
359 ca->shadow_wnd = tp->snd_cwnd;
360 break;
361 case CA_EVENT_COMPLETE_CWR:
362 ca->state = CDG_UNKNOWN;
363 ca->rtt_seq = tp->snd_nxt;
364 ca->rtt_prev = ca->rtt;
365 ca->rtt.v64 = 0;
366 break;
367 default:
368 break;
369 }
370}
371
372static void tcp_cdg_init(struct sock *sk)
373{
374 struct cdg *ca = inet_csk_ca(sk);
375 struct tcp_sock *tp = tcp_sk(sk);
376
377 /* We silently fall back to window = 1 if allocation fails. */
378 if (window > 1)
379 ca->gradients = kcalloc(window, sizeof(ca->gradients[0]),
380 GFP_NOWAIT | __GFP_NOWARN);
381 ca->rtt_seq = tp->snd_nxt;
382 ca->shadow_wnd = tp->snd_cwnd;
383}
384
385static void tcp_cdg_release(struct sock *sk)
386{
387 struct cdg *ca = inet_csk_ca(sk);
388
389 kfree(ca->gradients);
390}
391
392static struct tcp_congestion_ops tcp_cdg __read_mostly = {
393 .cong_avoid = tcp_cdg_cong_avoid,
394 .cwnd_event = tcp_cdg_cwnd_event,
395 .pkts_acked = tcp_cdg_acked,
396 .undo_cwnd = tcp_reno_undo_cwnd,
397 .ssthresh = tcp_cdg_ssthresh,
398 .release = tcp_cdg_release,
399 .init = tcp_cdg_init,
400 .owner = THIS_MODULE,
401 .name = "cdg",
402};
403
404static int __init tcp_cdg_register(void)
405{
406 if (backoff_beta > 1024 || window < 1 || window > 256)
407 return -ERANGE;
408 if (!is_power_of_2(window))
409 return -EINVAL;
410
411 BUILD_BUG_ON(sizeof(struct cdg) > ICSK_CA_PRIV_SIZE);
412 tcp_register_congestion_control(&tcp_cdg);
413 return 0;
414}
415
416static void __exit tcp_cdg_unregister(void)
417{
418 tcp_unregister_congestion_control(&tcp_cdg);
419}
420
421module_init(tcp_cdg_register);
422module_exit(tcp_cdg_unregister);
423MODULE_AUTHOR("Kenneth Klette Jonassen");
424MODULE_LICENSE("GPL");
425MODULE_DESCRIPTION("TCP CDG");
426