xref: /openbmc/linux/net/ipv4/tcp_cdg.c (revision 4f2c0a4acffbec01079c28f839422e64ddeff004)
1  // SPDX-License-Identifier: GPL-2.0-only
2  /*
3   * CAIA Delay-Gradient (CDG) congestion control
4   *
5   * This implementation is based on the paper:
6   *   D.A. Hayes and G. Armitage. "Revisiting TCP congestion control using
7   *   delay gradients." In IFIP Networking, pages 328-341. Springer, 2011.
8   *
9   * Scavenger traffic (Less-than-Best-Effort) should disable coexistence
10   * heuristics using parameters use_shadow=0 and use_ineff=0.
11   *
12   * Parameters window, backoff_beta, and backoff_factor are crucial for
13   * throughput and delay. Future work is needed to determine better defaults,
14   * and to provide guidelines for use in different environments/contexts.
15   *
16   * Except for window, knobs are configured via /sys/module/tcp_cdg/parameters/.
17   * Parameter window is only configurable when loading tcp_cdg as a module.
18   *
19   * Notable differences from paper/FreeBSD:
20   *   o Using Hybrid Slow start and Proportional Rate Reduction.
21   *   o Add toggle for shadow window mechanism. Suggested by David Hayes.
22   *   o Add toggle for non-congestion loss tolerance.
23   *   o Scaling parameter G is changed to a backoff factor;
24   *     conversion is given by: backoff_factor = 1000/(G * window).
25   *   o Limit shadow window to 2 * cwnd, or to cwnd when application limited.
26   *   o More accurate e^-x.
27   */
28  #include <linux/kernel.h>
29  #include <linux/random.h>
30  #include <linux/module.h>
31  #include <linux/sched/clock.h>
32  
33  #include <net/tcp.h>
34  
35  #define HYSTART_ACK_TRAIN	1
36  #define HYSTART_DELAY		2
37  
38  static int window __read_mostly = 8;
39  static unsigned int backoff_beta __read_mostly = 0.7071 * 1024; /* sqrt 0.5 */
40  static unsigned int backoff_factor __read_mostly = 42;
41  static unsigned int hystart_detect __read_mostly = 3;
42  static unsigned int use_ineff __read_mostly = 5;
43  static bool use_shadow __read_mostly = true;
44  static bool use_tolerance __read_mostly;
45  
46  module_param(window, int, 0444);
47  MODULE_PARM_DESC(window, "gradient window size (power of two <= 256)");
48  module_param(backoff_beta, uint, 0644);
49  MODULE_PARM_DESC(backoff_beta, "backoff beta (0-1024)");
50  module_param(backoff_factor, uint, 0644);
51  MODULE_PARM_DESC(backoff_factor, "backoff probability scale factor");
52  module_param(hystart_detect, uint, 0644);
53  MODULE_PARM_DESC(hystart_detect, "use Hybrid Slow start "
54  		 "(0: disabled, 1: ACK train, 2: delay threshold, 3: both)");
55  module_param(use_ineff, uint, 0644);
56  MODULE_PARM_DESC(use_ineff, "use ineffectual backoff detection (threshold)");
57  module_param(use_shadow, bool, 0644);
58  MODULE_PARM_DESC(use_shadow, "use shadow window heuristic");
59  module_param(use_tolerance, bool, 0644);
60  MODULE_PARM_DESC(use_tolerance, "use loss tolerance heuristic");
61  
62  struct cdg_minmax {
63  	union {
64  		struct {
65  			s32 min;
66  			s32 max;
67  		};
68  		u64 v64;
69  	};
70  };
71  
72  enum cdg_state {
73  	CDG_UNKNOWN = 0,
74  	CDG_NONFULL = 1,
75  	CDG_FULL    = 2,
76  	CDG_BACKOFF = 3,
77  };
78  
79  struct cdg {
80  	struct cdg_minmax rtt;
81  	struct cdg_minmax rtt_prev;
82  	struct cdg_minmax *gradients;
83  	struct cdg_minmax gsum;
84  	bool gfilled;
85  	u8  tail;
86  	u8  state;
87  	u8  delack;
88  	u32 rtt_seq;
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   */
nexp_u32(u32 ux)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   */
tcp_cdg_hystart_update(struct sock * sk)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 = tp->tcp_mstamp;
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  					      tcp_snd_cwnd(tp));
165  				tp->snd_ssthresh = tcp_snd_cwnd(tp);
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  					      tcp_snd_cwnd(tp));
184  				tp->snd_ssthresh = tcp_snd_cwnd(tp);
185  			}
186  		}
187  	}
188  }
189  
tcp_cdg_grad(struct cdg * ca)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  
tcp_cdg_backoff(struct sock * sk,u32 grad)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 (get_random_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, tcp_snd_cwnd(tp));
256  	ca->state = CDG_BACKOFF;
257  	tcp_enter_cwr(sk);
258  	return true;
259  }
260  
261  /* Not called in CWR or Recovery state. */
tcp_cdg_cong_avoid(struct sock * sk,u32 ack,u32 acked)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, tcp_snd_cwnd(tp));
289  		return;
290  	}
291  
292  	prior_snd_cwnd = tcp_snd_cwnd(tp);
293  	tcp_reno_cong_avoid(sk, ack, acked);
294  
295  	incr = tcp_snd_cwnd(tp) - prior_snd_cwnd;
296  	ca->shadow_wnd = max(ca->shadow_wnd, ca->shadow_wnd + incr);
297  }
298  
tcp_cdg_acked(struct sock * sk,const struct ack_sample * sample)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  
tcp_cdg_ssthresh(struct sock * sk)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  	if (ca->state == CDG_BACKOFF)
334  		return max(2U, (tcp_snd_cwnd(tp) * min(1024U, backoff_beta)) >> 10);
335  
336  	if (ca->state == CDG_NONFULL && use_tolerance)
337  		return tcp_snd_cwnd(tp);
338  
339  	ca->shadow_wnd = min(ca->shadow_wnd >> 1, tcp_snd_cwnd(tp));
340  	if (use_shadow)
341  		return max3(2U, ca->shadow_wnd, tcp_snd_cwnd(tp) >> 1);
342  	return max(2U, tcp_snd_cwnd(tp) >> 1);
343  }
344  
tcp_cdg_cwnd_event(struct sock * sk,const enum tcp_ca_event ev)345  static void tcp_cdg_cwnd_event(struct sock *sk, const enum tcp_ca_event ev)
346  {
347  	struct cdg *ca = inet_csk_ca(sk);
348  	struct tcp_sock *tp = tcp_sk(sk);
349  	struct cdg_minmax *gradients;
350  
351  	switch (ev) {
352  	case CA_EVENT_CWND_RESTART:
353  		gradients = ca->gradients;
354  		if (gradients)
355  			memset(gradients, 0, window * sizeof(gradients[0]));
356  		memset(ca, 0, sizeof(*ca));
357  
358  		ca->gradients = gradients;
359  		ca->rtt_seq = tp->snd_nxt;
360  		ca->shadow_wnd = tcp_snd_cwnd(tp);
361  		break;
362  	case CA_EVENT_COMPLETE_CWR:
363  		ca->state = CDG_UNKNOWN;
364  		ca->rtt_seq = tp->snd_nxt;
365  		ca->rtt_prev = ca->rtt;
366  		ca->rtt.v64 = 0;
367  		break;
368  	default:
369  		break;
370  	}
371  }
372  
tcp_cdg_init(struct sock * sk)373  static void tcp_cdg_init(struct sock *sk)
374  {
375  	struct cdg *ca = inet_csk_ca(sk);
376  	struct tcp_sock *tp = tcp_sk(sk);
377  
378  	ca->gradients = NULL;
379  	/* We silently fall back to window = 1 if allocation fails. */
380  	if (window > 1)
381  		ca->gradients = kcalloc(window, sizeof(ca->gradients[0]),
382  					GFP_NOWAIT | __GFP_NOWARN);
383  	ca->rtt_seq = tp->snd_nxt;
384  	ca->shadow_wnd = tcp_snd_cwnd(tp);
385  }
386  
tcp_cdg_release(struct sock * sk)387  static void tcp_cdg_release(struct sock *sk)
388  {
389  	struct cdg *ca = inet_csk_ca(sk);
390  
391  	kfree(ca->gradients);
392  	ca->gradients = NULL;
393  }
394  
395  static struct tcp_congestion_ops tcp_cdg __read_mostly = {
396  	.cong_avoid = tcp_cdg_cong_avoid,
397  	.cwnd_event = tcp_cdg_cwnd_event,
398  	.pkts_acked = tcp_cdg_acked,
399  	.undo_cwnd = tcp_reno_undo_cwnd,
400  	.ssthresh = tcp_cdg_ssthresh,
401  	.release = tcp_cdg_release,
402  	.init = tcp_cdg_init,
403  	.owner = THIS_MODULE,
404  	.name = "cdg",
405  };
406  
tcp_cdg_register(void)407  static int __init tcp_cdg_register(void)
408  {
409  	if (backoff_beta > 1024 || window < 1 || window > 256)
410  		return -ERANGE;
411  	if (!is_power_of_2(window))
412  		return -EINVAL;
413  
414  	BUILD_BUG_ON(sizeof(struct cdg) > ICSK_CA_PRIV_SIZE);
415  	tcp_register_congestion_control(&tcp_cdg);
416  	return 0;
417  }
418  
tcp_cdg_unregister(void)419  static void __exit tcp_cdg_unregister(void)
420  {
421  	tcp_unregister_congestion_control(&tcp_cdg);
422  }
423  
424  module_init(tcp_cdg_register);
425  module_exit(tcp_cdg_unregister);
426  MODULE_AUTHOR("Kenneth Klette Jonassen");
427  MODULE_LICENSE("GPL");
428  MODULE_DESCRIPTION("TCP CDG");
429