1/*
2 * TCP CUBIC: Binary Increase Congestion control for TCP v2.1
3 *
4 * This is from the implementation of CUBIC TCP in
5 * Injong Rhee, Lisong Xu.
6 *  "CUBIC: A New TCP-Friendly High-Speed TCP Variant
7 *  in PFLDnet 2005
8 * Available from:
9 *  http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/cubic-paper.pdf
10 *
11 * Unless CUBIC is enabled and congestion window is large
12 * this behaves the same as the original Reno.
13 */
14
15#include <linux/mm.h>
16#include <linux/module.h>
17#include <net/tcp.h>
18#include <asm/div64.h>
19
20#define BICTCP_BETA_SCALE    1024	/* Scale factor beta calculation
21					 * max_cwnd = snd_cwnd * beta
22					 */
23#define BICTCP_B		4	 /*
24					  * In binary search,
25					  * go to point (max+min)/N
26					  */
27#define	BICTCP_HZ		10	/* BIC HZ 2^10 = 1024 */
28
29static int fast_convergence __read_mostly = 1;
30static int max_increment __read_mostly = 16;
31static int beta __read_mostly = 819;	/* = 819/1024 (BICTCP_BETA_SCALE) */
32static int initial_ssthresh __read_mostly;
33static int bic_scale __read_mostly = 41;
34static int tcp_friendliness __read_mostly = 1;
35
36static u32 cube_rtt_scale __read_mostly;
37static u32 beta_scale __read_mostly;
38static u64 cube_factor __read_mostly;
39
40/* Note parameters that are used for precomputing scale factors are read-only */
41module_param(fast_convergence, int, 0644);
42MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
43module_param(max_increment, int, 0644);
44MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
45module_param(beta, int, 0444);
46MODULE_PARM_DESC(beta, "beta for multiplicative increase");
47module_param(initial_ssthresh, int, 0644);
48MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
49module_param(bic_scale, int, 0444);
50MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
51module_param(tcp_friendliness, int, 0644);
52MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
53
54/* BIC TCP Parameters */
55struct bictcp {
56	u32	cnt;		/* increase cwnd by 1 after ACKs */
57	u32 	last_max_cwnd;	/* last maximum snd_cwnd */
58	u32	loss_cwnd;	/* congestion window at last loss */
59	u32	last_cwnd;	/* the last snd_cwnd */
60	u32	last_time;	/* time when updated last_cwnd */
61	u32	bic_origin_point;/* origin point of bic function */
62	u32	bic_K;		/* time to origin point from the beginning of the current epoch */
63	u32	delay_min;	/* min delay */
64	u32	epoch_start;	/* beginning of an epoch */
65	u32	ack_cnt;	/* number of acks */
66	u32	tcp_cwnd;	/* estimated tcp cwnd */
67#define ACK_RATIO_SHIFT	4
68	u32	delayed_ack;	/* estimate the ratio of Packets/ACKs << 4 */
69};
70
71static inline void bictcp_reset(struct bictcp *ca)
72{
73	ca->cnt = 0;
74	ca->last_max_cwnd = 0;
75	ca->loss_cwnd = 0;
76	ca->last_cwnd = 0;
77	ca->last_time = 0;
78	ca->bic_origin_point = 0;
79	ca->bic_K = 0;
80	ca->delay_min = 0;
81	ca->epoch_start = 0;
82	ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
83	ca->ack_cnt = 0;
84	ca->tcp_cwnd = 0;
85}
86
87static void bictcp_init(struct sock *sk)
88{
89	bictcp_reset(inet_csk_ca(sk));
90	if (initial_ssthresh)
91		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
92}
93
94/* calculate the cubic root of x using a table lookup followed by one
95 * Newton-Raphson iteration.
96 * Avg err ~= 0.195%
97 */
98static u32 cubic_root(u64 a)
99{
100	u32 x, b, shift;
101	/*
102	 * cbrt(x) MSB values for x MSB values in [0..63].
103	 * Precomputed then refined by hand - Willy Tarreau
104	 *
105	 * For x in [0..63],
106	 *   v = cbrt(x << 18) - 1
107	 *   cbrt(x) = (v[x] + 10) >> 6
108	 */
109	static const u8 v[] = {
110		/* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
111		/* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
112		/* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
113		/* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
114		/* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
115		/* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
116		/* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
117		/* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
118	};
119
120	b = fls64(a);
121	if (b < 7) {
122		/* a in [0..63] */
123		return ((u32)v[(u32)a] + 35) >> 6;
124	}
125
126	b = ((b * 84) >> 8) - 1;
127	shift = (a >> (b * 3));
128
129	x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
130
131	/*
132	 * Newton-Raphson iteration
133	 *                         2
134	 * x    = ( 2 * x  +  a / x  ) / 3
135	 *  k+1          k         k
136	 */
137	x = (2 * x + (u32)div64_64(a, (u64)x * (u64)(x - 1)));
138	x = ((x * 341) >> 10);
139	return x;
140}
141
142/*
143 * Compute congestion window to use.
144 */
145static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
146{
147	u64 offs;
148	u32 delta, t, bic_target, min_cnt, max_cnt;
149
150	ca->ack_cnt++;	/* count the number of ACKs */
151
152	if (ca->last_cwnd == cwnd &&
153	    (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
154		return;
155
156	ca->last_cwnd = cwnd;
157	ca->last_time = tcp_time_stamp;
158
159	if (ca->epoch_start == 0) {
160		ca->epoch_start = tcp_time_stamp;	/* record the beginning of an epoch */
161		ca->ack_cnt = 1;			/* start counting */
162		ca->tcp_cwnd = cwnd;			/* syn with cubic */
163
164		if (ca->last_max_cwnd <= cwnd) {
165			ca->bic_K = 0;
166			ca->bic_origin_point = cwnd;
167		} else {
168			/* Compute new K based on
169			 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
170			 */
171			ca->bic_K = cubic_root(cube_factor
172					       * (ca->last_max_cwnd - cwnd));
173			ca->bic_origin_point = ca->last_max_cwnd;
174		}
175	}
176
177	/* cubic function - calc*/
178	/* calculate c * time^3 / rtt,
179	 *  while considering overflow in calculation of time^3
180	 * (so time^3 is done by using 64 bit)
181	 * and without the support of division of 64bit numbers
182	 * (so all divisions are done by using 32 bit)
183	 *  also NOTE the unit of those veriables
184	 *	  time  = (t - K) / 2^bictcp_HZ
185	 *	  c = bic_scale >> 10
186	 * rtt  = (srtt >> 3) / HZ
187	 * !!! The following code does not have overflow problems,
188	 * if the cwnd < 1 million packets !!!
189	 */
190
191	/* change the unit from HZ to bictcp_HZ */
192	t = ((tcp_time_stamp + (ca->delay_min>>3) - ca->epoch_start)
193	     << BICTCP_HZ) / HZ;
194
195	if (t < ca->bic_K)		/* t - K */
196		offs = ca->bic_K - t;
197	else
198		offs = t - ca->bic_K;
199
200	/* c/rtt * (t-K)^3 */
201	delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
202	if (t < ca->bic_K)                                	/* below origin*/
203		bic_target = ca->bic_origin_point - delta;
204	else                                                	/* above origin*/
205		bic_target = ca->bic_origin_point + delta;
206
207	/* cubic function - calc bictcp_cnt*/
208	if (bic_target > cwnd) {
209		ca->cnt = cwnd / (bic_target - cwnd);
210	} else {
211		ca->cnt = 100 * cwnd;              /* very small increment*/
212	}
213
214	if (ca->delay_min > 0) {
215		/* max increment = Smax * rtt / 0.1  */
216		min_cnt = (cwnd * HZ * 8)/(10 * max_increment * ca->delay_min);
217
218		/* use concave growth when the target is above the origin */
219		if (ca->cnt < min_cnt && t >= ca->bic_K)
220			ca->cnt = min_cnt;
221	}
222
223	/* slow start and low utilization  */
224	if (ca->loss_cwnd == 0)		/* could be aggressive in slow start */
225		ca->cnt = 50;
226
227	/* TCP Friendly */
228	if (tcp_friendliness) {
229		u32 scale = beta_scale;
230		delta = (cwnd * scale) >> 3;
231		while (ca->ack_cnt > delta) {		/* update tcp cwnd */
232			ca->ack_cnt -= delta;
233			ca->tcp_cwnd++;
234		}
235
236		if (ca->tcp_cwnd > cwnd){	/* if bic is slower than tcp */
237			delta = ca->tcp_cwnd - cwnd;
238			max_cnt = cwnd / delta;
239			if (ca->cnt > max_cnt)
240				ca->cnt = max_cnt;
241		}
242	}
243
244	ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
245	if (ca->cnt == 0)			/* cannot be zero */
246		ca->cnt = 1;
247}
248
249
250/* Keep track of minimum rtt */
251static inline void measure_delay(struct sock *sk)
252{
253	const struct tcp_sock *tp = tcp_sk(sk);
254	struct bictcp *ca = inet_csk_ca(sk);
255	u32 delay;
256
257	/* No time stamp */
258	if (!(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) ||
259	     /* Discard delay samples right after fast recovery */
260	    (s32)(tcp_time_stamp - ca->epoch_start) < HZ)
261		return;
262
263	delay = (tcp_time_stamp - tp->rx_opt.rcv_tsecr)<<3;
264	if (delay == 0)
265		delay = 1;
266
267	/* first time call or link delay decreases */
268	if (ca->delay_min == 0 || ca->delay_min > delay)
269		ca->delay_min = delay;
270}
271
272static void bictcp_cong_avoid(struct sock *sk, u32 ack,
273			      u32 seq_rtt, u32 in_flight, int data_acked)
274{
275	struct tcp_sock *tp = tcp_sk(sk);
276	struct bictcp *ca = inet_csk_ca(sk);
277
278	if (data_acked)
279		measure_delay(sk);
280
281	if (!tcp_is_cwnd_limited(sk, in_flight))
282		return;
283
284	if (tp->snd_cwnd <= tp->snd_ssthresh)
285		tcp_slow_start(tp);
286	else {
287		bictcp_update(ca, tp->snd_cwnd);
288
289		/* In dangerous area, increase slowly.
290		 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
291		 */
292		if (tp->snd_cwnd_cnt >= ca->cnt) {
293			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
294				tp->snd_cwnd++;
295			tp->snd_cwnd_cnt = 0;
296		} else
297			tp->snd_cwnd_cnt++;
298	}
299
300}
301
302static u32 bictcp_recalc_ssthresh(struct sock *sk)
303{
304	const struct tcp_sock *tp = tcp_sk(sk);
305	struct bictcp *ca = inet_csk_ca(sk);
306
307	ca->epoch_start = 0;	/* end of epoch */
308
309	/* Wmax and fast convergence */
310	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
311		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
312			/ (2 * BICTCP_BETA_SCALE);
313	else
314		ca->last_max_cwnd = tp->snd_cwnd;
315
316	ca->loss_cwnd = tp->snd_cwnd;
317
318	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
319}
320
321static u32 bictcp_undo_cwnd(struct sock *sk)
322{
323	struct bictcp *ca = inet_csk_ca(sk);
324
325	return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
326}
327
328static void bictcp_state(struct sock *sk, u8 new_state)
329{
330	if (new_state == TCP_CA_Loss)
331		bictcp_reset(inet_csk_ca(sk));
332}
333
334/* Track delayed acknowledgment ratio using sliding window
335 * ratio = (15*ratio + sample) / 16
336 */
337static void bictcp_acked(struct sock *sk, u32 cnt, ktime_t last)
338{
339	const struct inet_connection_sock *icsk = inet_csk(sk);
340
341	if (cnt > 0 && icsk->icsk_ca_state == TCP_CA_Open) {
342		struct bictcp *ca = inet_csk_ca(sk);
343		cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
344		ca->delayed_ack += cnt;
345	}
346}
347
348
349static struct tcp_congestion_ops cubictcp = {
350	.init		= bictcp_init,
351	.ssthresh	= bictcp_recalc_ssthresh,
352	.cong_avoid	= bictcp_cong_avoid,
353	.set_state	= bictcp_state,
354	.undo_cwnd	= bictcp_undo_cwnd,
355	.pkts_acked     = bictcp_acked,
356	.owner		= THIS_MODULE,
357	.name		= "cubic",
358};
359
360static int __init cubictcp_register(void)
361{
362	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
363
364	/* Precompute a bunch of the scaling factors that are used per-packet
365	 * based on SRTT of 100ms
366	 */
367
368	beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
369
370	cube_rtt_scale = (bic_scale * 10);	/* 1024*c/rtt */
371
372	/* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
373	 *  so K = cubic_root( (wmax-cwnd)*rtt/c )
374	 * the unit of K is bictcp_HZ=2^10, not HZ
375	 *
376	 *  c = bic_scale >> 10
377	 *  rtt = 100ms
378	 *
379	 * the following code has been designed and tested for
380	 * cwnd < 1 million packets
381	 * RTT < 100 seconds
382	 * HZ < 1,000,00  (corresponding to 10 nano-second)
383	 */
384
385	/* 1/c * 2^2*bictcp_HZ * srtt */
386	cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
387
388	/* divide by bic_scale and by constant Srtt (100ms) */
389	do_div(cube_factor, bic_scale * 10);
390
391	return tcp_register_congestion_control(&cubictcp);
392}
393
394static void __exit cubictcp_unregister(void)
395{
396	tcp_unregister_congestion_control(&cubictcp);
397}
398
399module_init(cubictcp_register);
400module_exit(cubictcp_unregister);
401
402MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
403MODULE_LICENSE("GPL");
404MODULE_DESCRIPTION("CUBIC TCP");
405MODULE_VERSION("2.1");
406