2 * TCP Illinois congestion control.
4 * http://www.ews.uiuc.edu/~shaoliu/tcpillinois/index.html
6 * The algorithm is described in:
7 * "TCP-Illinois: A Loss and Delay-Based Congestion Control Algorithm
8 * for High-Speed Networks"
9 * http://www.ews.uiuc.edu/~shaoliu/papersandslides/liubassri06perf.pdf
11 * Implemented from description in paper and ns-2 simulation.
12 * Copyright (C) 2007 Stephen Hemminger <shemminger@linux-foundation.org>
15 #include <linux/module.h>
16 #include <linux/skbuff.h>
17 #include <linux/inet_diag.h>
18 #include <asm/div64.h>
22 #define ALPHA_SCALE (1u<<ALPHA_SHIFT)
23 #define ALPHA_MIN ((3*ALPHA_SCALE)/10) /* ~0.3 */
24 #define ALPHA_MAX (10*ALPHA_SCALE) /* 10.0 */
25 #define ALPHA_BASE ALPHA_SCALE /* 1.0 */
26 #define U32_MAX ((u32)~0U)
27 #define RTT_MAX (U32_MAX / ALPHA_MAX) /* 3.3 secs */
30 #define BETA_SCALE (1u<<BETA_SHIFT)
31 #define BETA_MIN (BETA_SCALE/8) /* 0.125 */
32 #define BETA_MAX (BETA_SCALE/2) /* 0.5 */
33 #define BETA_BASE BETA_MAX
35 static int win_thresh __read_mostly = 15;
36 module_param(win_thresh, int, 0);
37 MODULE_PARM_DESC(win_thresh, "Window threshold for starting adaptive sizing");
39 static int theta __read_mostly = 5;
40 module_param(theta, int, 0);
41 MODULE_PARM_DESC(theta, "# of fast RTT's before full growth");
43 /* TCP Illinois Parameters */
45 u64 sum_rtt; /* sum of rtt's measured within last rtt */
46 u16 cnt_rtt; /* # of rtts measured within last rtt */
47 u32 base_rtt; /* min of all rtt in usec */
48 u32 max_rtt; /* max of all rtt in usec */
49 u32 end_seq; /* right edge of current RTT */
50 u32 alpha; /* Additive increase */
51 u32 beta; /* Muliplicative decrease */
52 u16 acked; /* # packets acked by current ACK */
53 u8 rtt_above; /* average rtt has gone above threshold */
54 u8 rtt_low; /* # of rtts measurements below threshold */
57 static void rtt_reset(struct sock *sk)
59 struct tcp_sock *tp = tcp_sk(sk);
60 struct illinois *ca = inet_csk_ca(sk);
62 ca->end_seq = tp->snd_nxt;
66 /* TODO: age max_rtt? */
69 static void tcp_illinois_init(struct sock *sk)
71 struct illinois *ca = inet_csk_ca(sk);
73 ca->alpha = ALPHA_MAX;
75 ca->base_rtt = 0x7fffffff;
85 /* Measure RTT for each ack. */
86 static void tcp_illinois_acked(struct sock *sk, u32 pkts_acked, ktime_t last)
88 struct illinois *ca = inet_csk_ca(sk);
91 ca->acked = pkts_acked;
93 if (ktime_equal(last, net_invalid_timestamp()))
96 rtt = ktime_to_us(net_timedelta(last));
98 /* ignore bogus values, this prevents wraparound in alpha math */
102 /* keep track of minimum RTT seen so far */
103 if (ca->base_rtt > rtt)
107 if (ca->max_rtt < rtt)
114 /* Maximum queuing delay */
115 static inline u32 max_delay(const struct illinois *ca)
117 return ca->max_rtt - ca->base_rtt;
120 /* Average queuing delay */
121 static inline u32 avg_delay(const struct illinois *ca)
125 do_div(t, ca->cnt_rtt);
126 return t - ca->base_rtt;
130 * Compute value of alpha used for additive increase.
131 * If small window then use 1.0, equivalent to Reno.
133 * For larger windows, adjust based on average delay.
134 * A. If average delay is at minimum (we are uncongested),
135 * then use large alpha (10.0) to increase faster.
136 * B. If average delay is at maximum (getting congested)
137 * then use small alpha (0.3)
139 * The result is a convex window growth curve.
141 static u32 alpha(struct illinois *ca, u32 da, u32 dm)
143 u32 d1 = dm / 100; /* Low threshold */
146 /* If never got out of low delay zone, then use max */
150 /* Wait for 5 good RTT's before allowing alpha to go alpha max.
151 * This prevents one good RTT from causing sudden window increase.
153 if (++ca->rtt_low < theta)
166 * (dm - d1) amin amax
167 * k1 = -------------------
171 * k2 = ---------------- - d1
181 return (dm * ALPHA_MAX) /
182 (dm + (da * (ALPHA_MAX - ALPHA_MIN)) / ALPHA_MIN);
186 * Beta used for multiplicative decrease.
187 * For small window sizes returns same value as Reno (0.5)
189 * If delay is small (10% of max) then beta = 1/8
190 * If delay is up to 80% of max then beta = 1/2
191 * In between is a linear function
193 static u32 beta(u32 da, u32 dm)
202 if (da >= d3 || d3 <= d2)
209 * k3 = -------------------
218 return (BETA_MIN * d3 - BETA_MAX * d2 + (BETA_MAX - BETA_MIN) * da)
222 /* Update alpha and beta values once per RTT */
223 static void update_params(struct sock *sk)
225 struct tcp_sock *tp = tcp_sk(sk);
226 struct illinois *ca = inet_csk_ca(sk);
228 if (tp->snd_cwnd < win_thresh) {
229 ca->alpha = ALPHA_BASE;
230 ca->beta = BETA_BASE;
231 } else if (ca->cnt_rtt > 0) {
232 u32 dm = max_delay(ca);
233 u32 da = avg_delay(ca);
235 ca->alpha = alpha(ca, da, dm);
236 ca->beta = beta(da, dm);
243 * In case of loss, reset to default values
245 static void tcp_illinois_state(struct sock *sk, u8 new_state)
247 struct illinois *ca = inet_csk_ca(sk);
249 if (new_state == TCP_CA_Loss) {
250 ca->alpha = ALPHA_BASE;
251 ca->beta = BETA_BASE;
259 * Increase window in response to successful acknowledgment.
261 static void tcp_illinois_cong_avoid(struct sock *sk, u32 ack,
262 u32 in_flight, int flag)
264 struct tcp_sock *tp = tcp_sk(sk);
265 struct illinois *ca = inet_csk_ca(sk);
267 if (after(ack, ca->end_seq))
270 /* RFC2861 only increase cwnd if fully utilized */
271 if (!tcp_is_cwnd_limited(sk, in_flight))
275 if (tp->snd_cwnd <= tp->snd_ssthresh)
281 /* snd_cwnd_cnt is # of packets since last cwnd increment */
282 tp->snd_cwnd_cnt += ca->acked;
285 /* This is close approximation of:
286 * tp->snd_cwnd += alpha/tp->snd_cwnd
288 delta = (tp->snd_cwnd_cnt * ca->alpha) >> ALPHA_SHIFT;
289 if (delta >= tp->snd_cwnd) {
290 tp->snd_cwnd = min(tp->snd_cwnd + delta / tp->snd_cwnd,
291 (u32) tp->snd_cwnd_clamp);
292 tp->snd_cwnd_cnt = 0;
297 static u32 tcp_illinois_ssthresh(struct sock *sk)
299 struct tcp_sock *tp = tcp_sk(sk);
300 struct illinois *ca = inet_csk_ca(sk);
302 /* Multiplicative decrease */
303 return max((tp->snd_cwnd * ca->beta) >> BETA_SHIFT, 2U);
307 /* Extract info for Tcp socket info provided via netlink. */
308 static void tcp_illinois_info(struct sock *sk, u32 ext,
311 const struct illinois *ca = inet_csk_ca(sk);
313 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
314 struct tcpvegas_info info = {
316 .tcpv_rttcnt = ca->cnt_rtt,
317 .tcpv_minrtt = ca->base_rtt,
321 do_div(t, ca->cnt_rtt);
324 nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
328 static struct tcp_congestion_ops tcp_illinois = {
329 .flags = TCP_CONG_RTT_STAMP,
330 .init = tcp_illinois_init,
331 .ssthresh = tcp_illinois_ssthresh,
332 .min_cwnd = tcp_reno_min_cwnd,
333 .cong_avoid = tcp_illinois_cong_avoid,
334 .set_state = tcp_illinois_state,
335 .get_info = tcp_illinois_info,
336 .pkts_acked = tcp_illinois_acked,
338 .owner = THIS_MODULE,
342 static int __init tcp_illinois_register(void)
344 BUILD_BUG_ON(sizeof(struct illinois) > ICSK_CA_PRIV_SIZE);
345 return tcp_register_congestion_control(&tcp_illinois);
348 static void __exit tcp_illinois_unregister(void)
350 tcp_unregister_congestion_control(&tcp_illinois);
353 module_init(tcp_illinois_register);
354 module_exit(tcp_illinois_unregister);
356 MODULE_AUTHOR("Stephen Hemminger, Shao Liu");
357 MODULE_LICENSE("GPL");
358 MODULE_DESCRIPTION("TCP Illinois");
359 MODULE_VERSION("1.0");