xref: /openbmc/linux/net/ipv4/tcp_cdg.c (revision 6dfcd296)
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 <net/tcp.h>
31 
32 #define HYSTART_ACK_TRAIN	1
33 #define HYSTART_DELAY		2
34 
35 static int window __read_mostly = 8;
36 static unsigned int backoff_beta __read_mostly = 0.7071 * 1024; /* sqrt 0.5 */
37 static unsigned int backoff_factor __read_mostly = 42;
38 static unsigned int hystart_detect __read_mostly = 3;
39 static unsigned int use_ineff __read_mostly = 5;
40 static bool use_shadow __read_mostly = true;
41 static bool use_tolerance __read_mostly;
42 
43 module_param(window, int, 0444);
44 MODULE_PARM_DESC(window, "gradient window size (power of two <= 256)");
45 module_param(backoff_beta, uint, 0644);
46 MODULE_PARM_DESC(backoff_beta, "backoff beta (0-1024)");
47 module_param(backoff_factor, uint, 0644);
48 MODULE_PARM_DESC(backoff_factor, "backoff probability scale factor");
49 module_param(hystart_detect, uint, 0644);
50 MODULE_PARM_DESC(hystart_detect, "use Hybrid Slow start "
51 		 "(0: disabled, 1: ACK train, 2: delay threshold, 3: both)");
52 module_param(use_ineff, uint, 0644);
53 MODULE_PARM_DESC(use_ineff, "use ineffectual backoff detection (threshold)");
54 module_param(use_shadow, bool, 0644);
55 MODULE_PARM_DESC(use_shadow, "use shadow window heuristic");
56 module_param(use_tolerance, bool, 0644);
57 MODULE_PARM_DESC(use_tolerance, "use loss tolerance heuristic");
58 
59 struct cdg_minmax {
60 	union {
61 		struct {
62 			s32 min;
63 			s32 max;
64 		};
65 		u64 v64;
66 	};
67 };
68 
69 enum cdg_state {
70 	CDG_UNKNOWN = 0,
71 	CDG_NONFULL = 1,
72 	CDG_FULL    = 2,
73 	CDG_BACKOFF = 3,
74 };
75 
76 struct cdg {
77 	struct cdg_minmax rtt;
78 	struct cdg_minmax rtt_prev;
79 	struct cdg_minmax *gradients;
80 	struct cdg_minmax gsum;
81 	bool gfilled;
82 	u8  tail;
83 	u8  state;
84 	u8  delack;
85 	u32 rtt_seq;
86 	u32 undo_cwnd;
87 	u32 shadow_wnd;
88 	u16 backoff_cnt;
89 	u16 sample_cnt;
90 	s32 delay_min;
91 	u32 last_ack;
92 	u32 round_start;
93 };
94 
95 /**
96  * nexp_u32 - negative base-e exponential
97  * @ux: x in units of micro
98  *
99  * Returns exp(ux * -1e-6) * U32_MAX.
100  */
101 static u32 __pure nexp_u32(u32 ux)
102 {
103 	static const u16 v[] = {
104 		/* exp(-x)*65536-1 for x = 0, 0.000256, 0.000512, ... */
105 		65535,
106 		65518, 65501, 65468, 65401, 65267, 65001, 64470, 63422,
107 		61378, 57484, 50423, 38795, 22965, 8047,  987,   14,
108 	};
109 	u32 msb = ux >> 8;
110 	u32 res;
111 	int i;
112 
113 	/* Cut off when ux >= 2^24 (actual result is <= 222/U32_MAX). */
114 	if (msb > U16_MAX)
115 		return 0;
116 
117 	/* Scale first eight bits linearly: */
118 	res = U32_MAX - (ux & 0xff) * (U32_MAX / 1000000);
119 
120 	/* Obtain e^(x + y + ...) by computing e^x * e^y * ...: */
121 	for (i = 1; msb; i++, msb >>= 1) {
122 		u32 y = v[i & -(msb & 1)] + U32_C(1);
123 
124 		res = ((u64)res * y) >> 16;
125 	}
126 
127 	return res;
128 }
129 
130 /* Based on the HyStart algorithm (by Ha et al.) that is implemented in
131  * tcp_cubic. Differences/experimental changes:
132  *   o Using Hayes' delayed ACK filter.
133  *   o Using a usec clock for the ACK train.
134  *   o Reset ACK train when application limited.
135  *   o Invoked at any cwnd (i.e. also when cwnd < 16).
136  *   o Invoked only when cwnd < ssthresh (i.e. not when cwnd == ssthresh).
137  */
138 static void tcp_cdg_hystart_update(struct sock *sk)
139 {
140 	struct cdg *ca = inet_csk_ca(sk);
141 	struct tcp_sock *tp = tcp_sk(sk);
142 
143 	ca->delay_min = min_not_zero(ca->delay_min, ca->rtt.min);
144 	if (ca->delay_min == 0)
145 		return;
146 
147 	if (hystart_detect & HYSTART_ACK_TRAIN) {
148 		u32 now_us = div_u64(local_clock(), NSEC_PER_USEC);
149 
150 		if (ca->last_ack == 0 || !tcp_is_cwnd_limited(sk)) {
151 			ca->last_ack = now_us;
152 			ca->round_start = now_us;
153 		} else if (before(now_us, ca->last_ack + 3000)) {
154 			u32 base_owd = max(ca->delay_min / 2U, 125U);
155 
156 			ca->last_ack = now_us;
157 			if (after(now_us, ca->round_start + base_owd)) {
158 				NET_INC_STATS(sock_net(sk),
159 					      LINUX_MIB_TCPHYSTARTTRAINDETECT);
160 				NET_ADD_STATS(sock_net(sk),
161 					      LINUX_MIB_TCPHYSTARTTRAINCWND,
162 					      tp->snd_cwnd);
163 				tp->snd_ssthresh = tp->snd_cwnd;
164 				return;
165 			}
166 		}
167 	}
168 
169 	if (hystart_detect & HYSTART_DELAY) {
170 		if (ca->sample_cnt < 8) {
171 			ca->sample_cnt++;
172 		} else {
173 			s32 thresh = max(ca->delay_min + ca->delay_min / 8U,
174 					 125U);
175 
176 			if (ca->rtt.min > thresh) {
177 				NET_INC_STATS(sock_net(sk),
178 					      LINUX_MIB_TCPHYSTARTDELAYDETECT);
179 				NET_ADD_STATS(sock_net(sk),
180 					      LINUX_MIB_TCPHYSTARTDELAYCWND,
181 					      tp->snd_cwnd);
182 				tp->snd_ssthresh = tp->snd_cwnd;
183 			}
184 		}
185 	}
186 }
187 
188 static s32 tcp_cdg_grad(struct cdg *ca)
189 {
190 	s32 gmin = ca->rtt.min - ca->rtt_prev.min;
191 	s32 gmax = ca->rtt.max - ca->rtt_prev.max;
192 	s32 grad;
193 
194 	if (ca->gradients) {
195 		ca->gsum.min += gmin - ca->gradients[ca->tail].min;
196 		ca->gsum.max += gmax - ca->gradients[ca->tail].max;
197 		ca->gradients[ca->tail].min = gmin;
198 		ca->gradients[ca->tail].max = gmax;
199 		ca->tail = (ca->tail + 1) & (window - 1);
200 		gmin = ca->gsum.min;
201 		gmax = ca->gsum.max;
202 	}
203 
204 	/* We keep sums to ignore gradients during cwnd reductions;
205 	 * the paper's smoothed gradients otherwise simplify to:
206 	 * (rtt_latest - rtt_oldest) / window.
207 	 *
208 	 * We also drop division by window here.
209 	 */
210 	grad = gmin > 0 ? gmin : gmax;
211 
212 	/* Extrapolate missing values in gradient window: */
213 	if (!ca->gfilled) {
214 		if (!ca->gradients && window > 1)
215 			grad *= window; /* Memory allocation failed. */
216 		else if (ca->tail == 0)
217 			ca->gfilled = true;
218 		else
219 			grad = (grad * window) / (int)ca->tail;
220 	}
221 
222 	/* Backoff was effectual: */
223 	if (gmin <= -32 || gmax <= -32)
224 		ca->backoff_cnt = 0;
225 
226 	if (use_tolerance) {
227 		/* Reduce small variations to zero: */
228 		gmin = DIV_ROUND_CLOSEST(gmin, 64);
229 		gmax = DIV_ROUND_CLOSEST(gmax, 64);
230 
231 		if (gmin > 0 && gmax <= 0)
232 			ca->state = CDG_FULL;
233 		else if ((gmin > 0 && gmax > 0) || gmax < 0)
234 			ca->state = CDG_NONFULL;
235 	}
236 	return grad;
237 }
238 
239 static bool tcp_cdg_backoff(struct sock *sk, u32 grad)
240 {
241 	struct cdg *ca = inet_csk_ca(sk);
242 	struct tcp_sock *tp = tcp_sk(sk);
243 
244 	if (prandom_u32() <= nexp_u32(grad * backoff_factor))
245 		return false;
246 
247 	if (use_ineff) {
248 		ca->backoff_cnt++;
249 		if (ca->backoff_cnt > use_ineff)
250 			return false;
251 	}
252 
253 	ca->shadow_wnd = max(ca->shadow_wnd, tp->snd_cwnd);
254 	ca->state = CDG_BACKOFF;
255 	tcp_enter_cwr(sk);
256 	return true;
257 }
258 
259 /* Not called in CWR or Recovery state. */
260 static void tcp_cdg_cong_avoid(struct sock *sk, u32 ack, u32 acked)
261 {
262 	struct cdg *ca = inet_csk_ca(sk);
263 	struct tcp_sock *tp = tcp_sk(sk);
264 	u32 prior_snd_cwnd;
265 	u32 incr;
266 
267 	if (tcp_in_slow_start(tp) && hystart_detect)
268 		tcp_cdg_hystart_update(sk);
269 
270 	if (after(ack, ca->rtt_seq) && ca->rtt.v64) {
271 		s32 grad = 0;
272 
273 		if (ca->rtt_prev.v64)
274 			grad = tcp_cdg_grad(ca);
275 		ca->rtt_seq = tp->snd_nxt;
276 		ca->rtt_prev = ca->rtt;
277 		ca->rtt.v64 = 0;
278 		ca->last_ack = 0;
279 		ca->sample_cnt = 0;
280 
281 		if (grad > 0 && tcp_cdg_backoff(sk, grad))
282 			return;
283 	}
284 
285 	if (!tcp_is_cwnd_limited(sk)) {
286 		ca->shadow_wnd = min(ca->shadow_wnd, tp->snd_cwnd);
287 		return;
288 	}
289 
290 	prior_snd_cwnd = tp->snd_cwnd;
291 	tcp_reno_cong_avoid(sk, ack, acked);
292 
293 	incr = tp->snd_cwnd - prior_snd_cwnd;
294 	ca->shadow_wnd = max(ca->shadow_wnd, ca->shadow_wnd + incr);
295 }
296 
297 static void tcp_cdg_acked(struct sock *sk, const struct ack_sample *sample)
298 {
299 	struct cdg *ca = inet_csk_ca(sk);
300 	struct tcp_sock *tp = tcp_sk(sk);
301 
302 	if (sample->rtt_us <= 0)
303 		return;
304 
305 	/* A heuristic for filtering delayed ACKs, adapted from:
306 	 * D.A. Hayes. "Timing enhancements to the FreeBSD kernel to support
307 	 * delay and rate based TCP mechanisms." TR 100219A. CAIA, 2010.
308 	 */
309 	if (tp->sacked_out == 0) {
310 		if (sample->pkts_acked == 1 && ca->delack) {
311 			/* A delayed ACK is only used for the minimum if it is
312 			 * provenly lower than an existing non-zero minimum.
313 			 */
314 			ca->rtt.min = min(ca->rtt.min, sample->rtt_us);
315 			ca->delack--;
316 			return;
317 		} else if (sample->pkts_acked > 1 && ca->delack < 5) {
318 			ca->delack++;
319 		}
320 	}
321 
322 	ca->rtt.min = min_not_zero(ca->rtt.min, sample->rtt_us);
323 	ca->rtt.max = max(ca->rtt.max, sample->rtt_us);
324 }
325 
326 static u32 tcp_cdg_ssthresh(struct sock *sk)
327 {
328 	struct cdg *ca = inet_csk_ca(sk);
329 	struct tcp_sock *tp = tcp_sk(sk);
330 
331 	ca->undo_cwnd = tp->snd_cwnd;
332 
333 	if (ca->state == CDG_BACKOFF)
334 		return max(2U, (tp->snd_cwnd * min(1024U, backoff_beta)) >> 10);
335 
336 	if (ca->state == CDG_NONFULL && use_tolerance)
337 		return tp->snd_cwnd;
338 
339 	ca->shadow_wnd = min(ca->shadow_wnd >> 1, tp->snd_cwnd);
340 	if (use_shadow)
341 		return max3(2U, ca->shadow_wnd, tp->snd_cwnd >> 1);
342 	return max(2U, tp->snd_cwnd >> 1);
343 }
344 
345 static u32 tcp_cdg_undo_cwnd(struct sock *sk)
346 {
347 	struct cdg *ca = inet_csk_ca(sk);
348 
349 	return max(tcp_sk(sk)->snd_cwnd, ca->undo_cwnd);
350 }
351 
352 static void tcp_cdg_cwnd_event(struct sock *sk, const enum tcp_ca_event ev)
353 {
354 	struct cdg *ca = inet_csk_ca(sk);
355 	struct tcp_sock *tp = tcp_sk(sk);
356 	struct cdg_minmax *gradients;
357 
358 	switch (ev) {
359 	case CA_EVENT_CWND_RESTART:
360 		gradients = ca->gradients;
361 		if (gradients)
362 			memset(gradients, 0, window * sizeof(gradients[0]));
363 		memset(ca, 0, sizeof(*ca));
364 
365 		ca->gradients = gradients;
366 		ca->rtt_seq = tp->snd_nxt;
367 		ca->shadow_wnd = tp->snd_cwnd;
368 		break;
369 	case CA_EVENT_COMPLETE_CWR:
370 		ca->state = CDG_UNKNOWN;
371 		ca->rtt_seq = tp->snd_nxt;
372 		ca->rtt_prev = ca->rtt;
373 		ca->rtt.v64 = 0;
374 		break;
375 	default:
376 		break;
377 	}
378 }
379 
380 static void tcp_cdg_init(struct sock *sk)
381 {
382 	struct cdg *ca = inet_csk_ca(sk);
383 	struct tcp_sock *tp = tcp_sk(sk);
384 
385 	/* We silently fall back to window = 1 if allocation fails. */
386 	if (window > 1)
387 		ca->gradients = kcalloc(window, sizeof(ca->gradients[0]),
388 					GFP_NOWAIT | __GFP_NOWARN);
389 	ca->rtt_seq = tp->snd_nxt;
390 	ca->shadow_wnd = tp->snd_cwnd;
391 }
392 
393 static void tcp_cdg_release(struct sock *sk)
394 {
395 	struct cdg *ca = inet_csk_ca(sk);
396 
397 	kfree(ca->gradients);
398 }
399 
400 struct tcp_congestion_ops tcp_cdg __read_mostly = {
401 	.cong_avoid = tcp_cdg_cong_avoid,
402 	.cwnd_event = tcp_cdg_cwnd_event,
403 	.pkts_acked = tcp_cdg_acked,
404 	.undo_cwnd = tcp_cdg_undo_cwnd,
405 	.ssthresh = tcp_cdg_ssthresh,
406 	.release = tcp_cdg_release,
407 	.init = tcp_cdg_init,
408 	.owner = THIS_MODULE,
409 	.name = "cdg",
410 };
411 
412 static int __init tcp_cdg_register(void)
413 {
414 	if (backoff_beta > 1024 || window < 1 || window > 256)
415 		return -ERANGE;
416 	if (!is_power_of_2(window))
417 		return -EINVAL;
418 
419 	BUILD_BUG_ON(sizeof(struct cdg) > ICSK_CA_PRIV_SIZE);
420 	tcp_register_congestion_control(&tcp_cdg);
421 	return 0;
422 }
423 
424 static void __exit tcp_cdg_unregister(void)
425 {
426 	tcp_unregister_congestion_control(&tcp_cdg);
427 }
428 
429 module_init(tcp_cdg_register);
430 module_exit(tcp_cdg_unregister);
431 MODULE_AUTHOR("Kenneth Klette Jonassen");
432 MODULE_LICENSE("GPL");
433 MODULE_DESCRIPTION("TCP CDG");
434