2 * H-TCP congestion control. The algorithm is detailed in:
3 * R.N.Shorten, D.J.Leith:
4 * "H-TCP: TCP for high-speed and long-distance networks"
5 * Proc. PFLDnet, Argonne, 2004.
6 * http://www.hamilton.ie/net/htcp3.pdf
10 #include <linux/module.h>
13 #define ALPHA_BASE (1<<7) /* 1.0 with shift << 7 */
14 #define BETA_MIN (1<<6) /* 0.5 with shift << 7 */
15 #define BETA_MAX 102 /* 0.8 with shift << 7 */
17 static int use_rtt_scaling __read_mostly = 1;
18 module_param(use_rtt_scaling, int, 0644);
19 MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling");
21 static int use_bandwidth_switch __read_mostly = 1;
22 module_param(use_bandwidth_switch, int, 0644);
23 MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher");
26 u32 alpha; /* Fixed point arith, << 7 */
27 u8 beta; /* Fixed point arith, << 7 */
28 u8 modeswitch; /* Delay modeswitch
29 until we had at least one congestion event */
34 u32 last_cong; /* Time since last congestion event end */
40 /* Bandwidth estimation */
48 static inline u32 htcp_cong_time(const struct htcp *ca)
50 return jiffies - ca->last_cong;
53 static inline u32 htcp_ccount(const struct htcp *ca)
55 return htcp_cong_time(ca) / ca->minRTT;
58 static inline void htcp_reset(struct htcp *ca)
60 ca->undo_last_cong = ca->last_cong;
61 ca->undo_maxRTT = ca->maxRTT;
62 ca->undo_old_maxB = ca->old_maxB;
64 ca->last_cong = jiffies;
67 static u32 htcp_cwnd_undo(struct sock *sk)
69 const struct tcp_sock *tp = tcp_sk(sk);
70 struct htcp *ca = inet_csk_ca(sk);
72 if (ca->undo_last_cong) {
73 ca->last_cong = ca->undo_last_cong;
74 ca->maxRTT = ca->undo_maxRTT;
75 ca->old_maxB = ca->undo_old_maxB;
76 ca->undo_last_cong = 0;
79 return max(tp->snd_cwnd, (tp->snd_ssthresh << 7) / ca->beta);
82 static inline void measure_rtt(struct sock *sk, u32 srtt)
84 const struct inet_connection_sock *icsk = inet_csk(sk);
85 struct htcp *ca = inet_csk_ca(sk);
87 /* keep track of minimum RTT seen so far, minRTT is zero at first */
88 if (ca->minRTT > srtt || !ca->minRTT)
92 if (icsk->icsk_ca_state == TCP_CA_Open) {
93 if (ca->maxRTT < ca->minRTT)
94 ca->maxRTT = ca->minRTT;
95 if (ca->maxRTT < srtt &&
96 srtt <= ca->maxRTT + msecs_to_jiffies(20))
101 static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked, s32 rtt)
103 const struct inet_connection_sock *icsk = inet_csk(sk);
104 const struct tcp_sock *tp = tcp_sk(sk);
105 struct htcp *ca = inet_csk_ca(sk);
106 u32 now = tcp_time_stamp;
108 if (icsk->icsk_ca_state == TCP_CA_Open)
109 ca->pkts_acked = pkts_acked;
112 measure_rtt(sk, usecs_to_jiffies(rtt));
114 if (!use_bandwidth_switch)
117 /* achieved throughput calculations */
118 if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_Disorder))) {
124 ca->packetcount += pkts_acked;
126 if (ca->packetcount >= tp->snd_cwnd - (ca->alpha >> 7 ? : 1) &&
127 now - ca->lasttime >= ca->minRTT &&
129 __u32 cur_Bi = ca->packetcount * HZ / (now - ca->lasttime);
131 if (htcp_ccount(ca) <= 3) {
132 /* just after backoff */
133 ca->minB = ca->maxB = ca->Bi = cur_Bi;
135 ca->Bi = (3 * ca->Bi + cur_Bi) / 4;
136 if (ca->Bi > ca->maxB)
138 if (ca->minB > ca->maxB)
146 static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
148 if (use_bandwidth_switch) {
150 u32 old_maxB = ca->old_maxB;
151 ca->old_maxB = ca->maxB;
153 if (!between(5 * maxB, 4 * old_maxB, 6 * old_maxB)) {
160 if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
161 ca->beta = (minRTT << 7) / maxRTT;
162 if (ca->beta < BETA_MIN)
164 else if (ca->beta > BETA_MAX)
172 static inline void htcp_alpha_update(struct htcp *ca)
174 u32 minRTT = ca->minRTT;
176 u32 diff = htcp_cong_time(ca);
180 factor = 1 + (10 * diff + ((diff / 2) * (diff / 2) / HZ)) / HZ;
183 if (use_rtt_scaling && minRTT) {
184 u32 scale = (HZ << 3) / (10 * minRTT);
186 /* clamping ratio to interval [0.5,10]<<3 */
187 scale = min(max(scale, 1U << 2), 10U << 3);
188 factor = (factor << 3) / scale;
193 ca->alpha = 2 * factor * ((1 << 7) - ca->beta);
195 ca->alpha = ALPHA_BASE;
199 * After we have the rtt data to calculate beta, we'd still prefer to wait one
200 * rtt before we adjust our beta to ensure we are working from a consistent
203 * This function should be called when we hit a congestion event since only at
204 * that point do we really have a real sense of maxRTT (the queues en route
205 * were getting just too full now).
207 static void htcp_param_update(struct sock *sk)
209 struct htcp *ca = inet_csk_ca(sk);
210 u32 minRTT = ca->minRTT;
211 u32 maxRTT = ca->maxRTT;
213 htcp_beta_update(ca, minRTT, maxRTT);
214 htcp_alpha_update(ca);
216 /* add slowly fading memory for maxRTT to accommodate routing changes */
217 if (minRTT > 0 && maxRTT > minRTT)
218 ca->maxRTT = minRTT + ((maxRTT - minRTT) * 95) / 100;
221 static u32 htcp_recalc_ssthresh(struct sock *sk)
223 const struct tcp_sock *tp = tcp_sk(sk);
224 const struct htcp *ca = inet_csk_ca(sk);
226 htcp_param_update(sk);
227 return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
230 static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
232 struct tcp_sock *tp = tcp_sk(sk);
233 struct htcp *ca = inet_csk_ca(sk);
235 if (!tcp_is_cwnd_limited(sk, in_flight))
238 if (tp->snd_cwnd <= tp->snd_ssthresh)
239 tcp_slow_start(tp, acked);
241 /* In dangerous area, increase slowly.
242 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
244 if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
245 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
247 tp->snd_cwnd_cnt = 0;
248 htcp_alpha_update(ca);
250 tp->snd_cwnd_cnt += ca->pkts_acked;
256 static void htcp_init(struct sock *sk)
258 struct htcp *ca = inet_csk_ca(sk);
260 memset(ca, 0, sizeof(struct htcp));
261 ca->alpha = ALPHA_BASE;
264 ca->last_cong = jiffies;
267 static void htcp_state(struct sock *sk, u8 new_state)
272 struct htcp *ca = inet_csk_ca(sk);
273 if (ca->undo_last_cong) {
274 ca->last_cong = jiffies;
275 ca->undo_last_cong = 0;
280 case TCP_CA_Recovery:
282 htcp_reset(inet_csk_ca(sk));
287 static struct tcp_congestion_ops htcp __read_mostly = {
289 .ssthresh = htcp_recalc_ssthresh,
290 .cong_avoid = htcp_cong_avoid,
291 .set_state = htcp_state,
292 .undo_cwnd = htcp_cwnd_undo,
293 .pkts_acked = measure_achieved_throughput,
294 .owner = THIS_MODULE,
298 static int __init htcp_register(void)
300 BUILD_BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
301 BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
302 return tcp_register_congestion_control(&htcp);
305 static void __exit htcp_unregister(void)
307 tcp_unregister_congestion_control(&htcp);
310 module_init(htcp_register);
311 module_exit(htcp_unregister);
313 MODULE_AUTHOR("Baruch Even");
314 MODULE_LICENSE("GPL");
315 MODULE_DESCRIPTION("H-TCP");