xref: /openbmc/linux/net/ipv4/tcp_cdg.c (revision 5d0e4d78)
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 
37 static int window __read_mostly = 8;
38 static unsigned int backoff_beta __read_mostly = 0.7071 * 1024; /* sqrt 0.5 */
39 static unsigned int backoff_factor __read_mostly = 42;
40 static unsigned int hystart_detect __read_mostly = 3;
41 static unsigned int use_ineff __read_mostly = 5;
42 static bool use_shadow __read_mostly = true;
43 static bool use_tolerance __read_mostly;
44 
45 module_param(window, int, 0444);
46 MODULE_PARM_DESC(window, "gradient window size (power of two <= 256)");
47 module_param(backoff_beta, uint, 0644);
48 MODULE_PARM_DESC(backoff_beta, "backoff beta (0-1024)");
49 module_param(backoff_factor, uint, 0644);
50 MODULE_PARM_DESC(backoff_factor, "backoff probability scale factor");
51 module_param(hystart_detect, uint, 0644);
52 MODULE_PARM_DESC(hystart_detect, "use Hybrid Slow start "
53 		 "(0: disabled, 1: ACK train, 2: delay threshold, 3: both)");
54 module_param(use_ineff, uint, 0644);
55 MODULE_PARM_DESC(use_ineff, "use ineffectual backoff detection (threshold)");
56 module_param(use_shadow, bool, 0644);
57 MODULE_PARM_DESC(use_shadow, "use shadow window heuristic");
58 module_param(use_tolerance, bool, 0644);
59 MODULE_PARM_DESC(use_tolerance, "use loss tolerance heuristic");
60 
61 struct cdg_minmax {
62 	union {
63 		struct {
64 			s32 min;
65 			s32 max;
66 		};
67 		u64 v64;
68 	};
69 };
70 
71 enum cdg_state {
72 	CDG_UNKNOWN = 0,
73 	CDG_NONFULL = 1,
74 	CDG_FULL    = 2,
75 	CDG_BACKOFF = 3,
76 };
77 
78 struct 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 undo_cwnd;
89 	u32 shadow_wnd;
90 	u16 backoff_cnt;
91 	u16 sample_cnt;
92 	s32 delay_min;
93 	u32 last_ack;
94 	u32 round_start;
95 };
96 
97 /**
98  * nexp_u32 - negative base-e exponential
99  * @ux: x in units of micro
100  *
101  * Returns exp(ux * -1e-6) * U32_MAX.
102  */
103 static u32 __pure nexp_u32(u32 ux)
104 {
105 	static const u16 v[] = {
106 		/* exp(-x)*65536-1 for x = 0, 0.000256, 0.000512, ... */
107 		65535,
108 		65518, 65501, 65468, 65401, 65267, 65001, 64470, 63422,
109 		61378, 57484, 50423, 38795, 22965, 8047,  987,   14,
110 	};
111 	u32 msb = ux >> 8;
112 	u32 res;
113 	int i;
114 
115 	/* Cut off when ux >= 2^24 (actual result is <= 222/U32_MAX). */
116 	if (msb > U16_MAX)
117 		return 0;
118 
119 	/* Scale first eight bits linearly: */
120 	res = U32_MAX - (ux & 0xff) * (U32_MAX / 1000000);
121 
122 	/* Obtain e^(x + y + ...) by computing e^x * e^y * ...: */
123 	for (i = 1; msb; i++, msb >>= 1) {
124 		u32 y = v[i & -(msb & 1)] + U32_C(1);
125 
126 		res = ((u64)res * y) >> 16;
127 	}
128 
129 	return res;
130 }
131 
132 /* Based on the HyStart algorithm (by Ha et al.) that is implemented in
133  * tcp_cubic. Differences/experimental changes:
134  *   o Using Hayes' delayed ACK filter.
135  *   o Using a usec clock for the ACK train.
136  *   o Reset ACK train when application limited.
137  *   o Invoked at any cwnd (i.e. also when cwnd < 16).
138  *   o Invoked only when cwnd < ssthresh (i.e. not when cwnd == ssthresh).
139  */
140 static void tcp_cdg_hystart_update(struct sock *sk)
141 {
142 	struct cdg *ca = inet_csk_ca(sk);
143 	struct tcp_sock *tp = tcp_sk(sk);
144 
145 	ca->delay_min = min_not_zero(ca->delay_min, ca->rtt.min);
146 	if (ca->delay_min == 0)
147 		return;
148 
149 	if (hystart_detect & HYSTART_ACK_TRAIN) {
150 		u32 now_us = div_u64(local_clock(), NSEC_PER_USEC);
151 
152 		if (ca->last_ack == 0 || !tcp_is_cwnd_limited(sk)) {
153 			ca->last_ack = now_us;
154 			ca->round_start = now_us;
155 		} else if (before(now_us, ca->last_ack + 3000)) {
156 			u32 base_owd = max(ca->delay_min / 2U, 125U);
157 
158 			ca->last_ack = now_us;
159 			if (after(now_us, ca->round_start + base_owd)) {
160 				NET_INC_STATS(sock_net(sk),
161 					      LINUX_MIB_TCPHYSTARTTRAINDETECT);
162 				NET_ADD_STATS(sock_net(sk),
163 					      LINUX_MIB_TCPHYSTARTTRAINCWND,
164 					      tp->snd_cwnd);
165 				tp->snd_ssthresh = tp->snd_cwnd;
166 				return;
167 			}
168 		}
169 	}
170 
171 	if (hystart_detect & HYSTART_DELAY) {
172 		if (ca->sample_cnt < 8) {
173 			ca->sample_cnt++;
174 		} else {
175 			s32 thresh = max(ca->delay_min + ca->delay_min / 8U,
176 					 125U);
177 
178 			if (ca->rtt.min > thresh) {
179 				NET_INC_STATS(sock_net(sk),
180 					      LINUX_MIB_TCPHYSTARTDELAYDETECT);
181 				NET_ADD_STATS(sock_net(sk),
182 					      LINUX_MIB_TCPHYSTARTDELAYCWND,
183 					      tp->snd_cwnd);
184 				tp->snd_ssthresh = tp->snd_cwnd;
185 			}
186 		}
187 	}
188 }
189 
190 static s32 tcp_cdg_grad(struct cdg *ca)
191 {
192 	s32 gmin = ca->rtt.min - ca->rtt_prev.min;
193 	s32 gmax = ca->rtt.max - ca->rtt_prev.max;
194 	s32 grad;
195 
196 	if (ca->gradients) {
197 		ca->gsum.min += gmin - ca->gradients[ca->tail].min;
198 		ca->gsum.max += gmax - ca->gradients[ca->tail].max;
199 		ca->gradients[ca->tail].min = gmin;
200 		ca->gradients[ca->tail].max = gmax;
201 		ca->tail = (ca->tail + 1) & (window - 1);
202 		gmin = ca->gsum.min;
203 		gmax = ca->gsum.max;
204 	}
205 
206 	/* We keep sums to ignore gradients during cwnd reductions;
207 	 * the paper's smoothed gradients otherwise simplify to:
208 	 * (rtt_latest - rtt_oldest) / window.
209 	 *
210 	 * We also drop division by window here.
211 	 */
212 	grad = gmin > 0 ? gmin : gmax;
213 
214 	/* Extrapolate missing values in gradient window: */
215 	if (!ca->gfilled) {
216 		if (!ca->gradients && window > 1)
217 			grad *= window; /* Memory allocation failed. */
218 		else if (ca->tail == 0)
219 			ca->gfilled = true;
220 		else
221 			grad = (grad * window) / (int)ca->tail;
222 	}
223 
224 	/* Backoff was effectual: */
225 	if (gmin <= -32 || gmax <= -32)
226 		ca->backoff_cnt = 0;
227 
228 	if (use_tolerance) {
229 		/* Reduce small variations to zero: */
230 		gmin = DIV_ROUND_CLOSEST(gmin, 64);
231 		gmax = DIV_ROUND_CLOSEST(gmax, 64);
232 
233 		if (gmin > 0 && gmax <= 0)
234 			ca->state = CDG_FULL;
235 		else if ((gmin > 0 && gmax > 0) || gmax < 0)
236 			ca->state = CDG_NONFULL;
237 	}
238 	return grad;
239 }
240 
241 static bool tcp_cdg_backoff(struct sock *sk, u32 grad)
242 {
243 	struct cdg *ca = inet_csk_ca(sk);
244 	struct tcp_sock *tp = tcp_sk(sk);
245 
246 	if (prandom_u32() <= nexp_u32(grad * backoff_factor))
247 		return false;
248 
249 	if (use_ineff) {
250 		ca->backoff_cnt++;
251 		if (ca->backoff_cnt > use_ineff)
252 			return false;
253 	}
254 
255 	ca->shadow_wnd = max(ca->shadow_wnd, tp->snd_cwnd);
256 	ca->state = CDG_BACKOFF;
257 	tcp_enter_cwr(sk);
258 	return true;
259 }
260 
261 /* Not called in CWR or Recovery state. */
262 static void tcp_cdg_cong_avoid(struct sock *sk, u32 ack, u32 acked)
263 {
264 	struct cdg *ca = inet_csk_ca(sk);
265 	struct tcp_sock *tp = tcp_sk(sk);
266 	u32 prior_snd_cwnd;
267 	u32 incr;
268 
269 	if (tcp_in_slow_start(tp) && hystart_detect)
270 		tcp_cdg_hystart_update(sk);
271 
272 	if (after(ack, ca->rtt_seq) && ca->rtt.v64) {
273 		s32 grad = 0;
274 
275 		if (ca->rtt_prev.v64)
276 			grad = tcp_cdg_grad(ca);
277 		ca->rtt_seq = tp->snd_nxt;
278 		ca->rtt_prev = ca->rtt;
279 		ca->rtt.v64 = 0;
280 		ca->last_ack = 0;
281 		ca->sample_cnt = 0;
282 
283 		if (grad > 0 && tcp_cdg_backoff(sk, grad))
284 			return;
285 	}
286 
287 	if (!tcp_is_cwnd_limited(sk)) {
288 		ca->shadow_wnd = min(ca->shadow_wnd, tp->snd_cwnd);
289 		return;
290 	}
291 
292 	prior_snd_cwnd = tp->snd_cwnd;
293 	tcp_reno_cong_avoid(sk, ack, acked);
294 
295 	incr = tp->snd_cwnd - prior_snd_cwnd;
296 	ca->shadow_wnd = max(ca->shadow_wnd, ca->shadow_wnd + incr);
297 }
298 
299 static void tcp_cdg_acked(struct sock *sk, const struct ack_sample *sample)
300 {
301 	struct cdg *ca = inet_csk_ca(sk);
302 	struct tcp_sock *tp = tcp_sk(sk);
303 
304 	if (sample->rtt_us <= 0)
305 		return;
306 
307 	/* A heuristic for filtering delayed ACKs, adapted from:
308 	 * D.A. Hayes. "Timing enhancements to the FreeBSD kernel to support
309 	 * delay and rate based TCP mechanisms." TR 100219A. CAIA, 2010.
310 	 */
311 	if (tp->sacked_out == 0) {
312 		if (sample->pkts_acked == 1 && ca->delack) {
313 			/* A delayed ACK is only used for the minimum if it is
314 			 * provenly lower than an existing non-zero minimum.
315 			 */
316 			ca->rtt.min = min(ca->rtt.min, sample->rtt_us);
317 			ca->delack--;
318 			return;
319 		} else if (sample->pkts_acked > 1 && ca->delack < 5) {
320 			ca->delack++;
321 		}
322 	}
323 
324 	ca->rtt.min = min_not_zero(ca->rtt.min, sample->rtt_us);
325 	ca->rtt.max = max(ca->rtt.max, sample->rtt_us);
326 }
327 
328 static u32 tcp_cdg_ssthresh(struct sock *sk)
329 {
330 	struct cdg *ca = inet_csk_ca(sk);
331 	struct tcp_sock *tp = tcp_sk(sk);
332 
333 	ca->undo_cwnd = tp->snd_cwnd;
334 
335 	if (ca->state == CDG_BACKOFF)
336 		return max(2U, (tp->snd_cwnd * min(1024U, backoff_beta)) >> 10);
337 
338 	if (ca->state == CDG_NONFULL && use_tolerance)
339 		return tp->snd_cwnd;
340 
341 	ca->shadow_wnd = min(ca->shadow_wnd >> 1, tp->snd_cwnd);
342 	if (use_shadow)
343 		return max3(2U, ca->shadow_wnd, tp->snd_cwnd >> 1);
344 	return max(2U, tp->snd_cwnd >> 1);
345 }
346 
347 static u32 tcp_cdg_undo_cwnd(struct sock *sk)
348 {
349 	struct cdg *ca = inet_csk_ca(sk);
350 
351 	return max(tcp_sk(sk)->snd_cwnd, ca->undo_cwnd);
352 }
353 
354 static void tcp_cdg_cwnd_event(struct sock *sk, const enum tcp_ca_event ev)
355 {
356 	struct cdg *ca = inet_csk_ca(sk);
357 	struct tcp_sock *tp = tcp_sk(sk);
358 	struct cdg_minmax *gradients;
359 
360 	switch (ev) {
361 	case CA_EVENT_CWND_RESTART:
362 		gradients = ca->gradients;
363 		if (gradients)
364 			memset(gradients, 0, window * sizeof(gradients[0]));
365 		memset(ca, 0, sizeof(*ca));
366 
367 		ca->gradients = gradients;
368 		ca->rtt_seq = tp->snd_nxt;
369 		ca->shadow_wnd = tp->snd_cwnd;
370 		break;
371 	case CA_EVENT_COMPLETE_CWR:
372 		ca->state = CDG_UNKNOWN;
373 		ca->rtt_seq = tp->snd_nxt;
374 		ca->rtt_prev = ca->rtt;
375 		ca->rtt.v64 = 0;
376 		break;
377 	default:
378 		break;
379 	}
380 }
381 
382 static void tcp_cdg_init(struct sock *sk)
383 {
384 	struct cdg *ca = inet_csk_ca(sk);
385 	struct tcp_sock *tp = tcp_sk(sk);
386 
387 	/* We silently fall back to window = 1 if allocation fails. */
388 	if (window > 1)
389 		ca->gradients = kcalloc(window, sizeof(ca->gradients[0]),
390 					GFP_NOWAIT | __GFP_NOWARN);
391 	ca->rtt_seq = tp->snd_nxt;
392 	ca->shadow_wnd = tp->snd_cwnd;
393 }
394 
395 static void tcp_cdg_release(struct sock *sk)
396 {
397 	struct cdg *ca = inet_csk_ca(sk);
398 
399 	kfree(ca->gradients);
400 }
401 
402 struct tcp_congestion_ops tcp_cdg __read_mostly = {
403 	.cong_avoid = tcp_cdg_cong_avoid,
404 	.cwnd_event = tcp_cdg_cwnd_event,
405 	.pkts_acked = tcp_cdg_acked,
406 	.undo_cwnd = tcp_cdg_undo_cwnd,
407 	.ssthresh = tcp_cdg_ssthresh,
408 	.release = tcp_cdg_release,
409 	.init = tcp_cdg_init,
410 	.owner = THIS_MODULE,
411 	.name = "cdg",
412 };
413 
414 static int __init tcp_cdg_register(void)
415 {
416 	if (backoff_beta > 1024 || window < 1 || window > 256)
417 		return -ERANGE;
418 	if (!is_power_of_2(window))
419 		return -EINVAL;
420 
421 	BUILD_BUG_ON(sizeof(struct cdg) > ICSK_CA_PRIV_SIZE);
422 	tcp_register_congestion_control(&tcp_cdg);
423 	return 0;
424 }
425 
426 static void __exit tcp_cdg_unregister(void)
427 {
428 	tcp_unregister_congestion_control(&tcp_cdg);
429 }
430 
431 module_init(tcp_cdg_register);
432 module_exit(tcp_cdg_unregister);
433 MODULE_AUTHOR("Kenneth Klette Jonassen");
434 MODULE_LICENSE("GPL");
435 MODULE_DESCRIPTION("TCP CDG");
436