sysctl mqueue: remove the binary sysctl numbers
[linux-2.6/kmemtrace.git] / net / ipv4 / tcp_htcp.c
blob5215691f2760e710ac56934e539665f51161fbe2
1 /*
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
7 */
9 #include <linux/mm.h>
10 #include <linux/module.h>
11 #include <net/tcp.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");
25 struct htcp {
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 */
30 u16 pkts_acked;
31 u32 packetcount;
32 u32 minRTT;
33 u32 maxRTT;
34 u32 last_cong; /* Time since last congestion event end */
35 u32 undo_last_cong;
37 u32 undo_maxRTT;
38 u32 undo_old_maxB;
40 /* Bandwidth estimation */
41 u32 minB;
42 u32 maxB;
43 u32 old_maxB;
44 u32 Bi;
45 u32 lasttime;
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 ca->last_cong = ca->undo_last_cong;
73 ca->maxRTT = ca->undo_maxRTT;
74 ca->old_maxB = ca->undo_old_maxB;
76 return max(tp->snd_cwnd, (tp->snd_ssthresh << 7) / ca->beta);
79 static inline void measure_rtt(struct sock *sk, u32 srtt)
81 const struct inet_connection_sock *icsk = inet_csk(sk);
82 struct htcp *ca = inet_csk_ca(sk);
84 /* keep track of minimum RTT seen so far, minRTT is zero at first */
85 if (ca->minRTT > srtt || !ca->minRTT)
86 ca->minRTT = srtt;
88 /* max RTT */
89 if (icsk->icsk_ca_state == TCP_CA_Open) {
90 if (ca->maxRTT < ca->minRTT)
91 ca->maxRTT = ca->minRTT;
92 if (ca->maxRTT < srtt
93 && srtt <= ca->maxRTT + msecs_to_jiffies(20))
94 ca->maxRTT = srtt;
98 static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked, s32 rtt)
100 const struct inet_connection_sock *icsk = inet_csk(sk);
101 const struct tcp_sock *tp = tcp_sk(sk);
102 struct htcp *ca = inet_csk_ca(sk);
103 u32 now = tcp_time_stamp;
105 if (icsk->icsk_ca_state == TCP_CA_Open)
106 ca->pkts_acked = pkts_acked;
108 if (rtt > 0)
109 measure_rtt(sk, usecs_to_jiffies(rtt));
111 if (!use_bandwidth_switch)
112 return;
114 /* achieved throughput calculations */
115 if (icsk->icsk_ca_state != TCP_CA_Open &&
116 icsk->icsk_ca_state != TCP_CA_Disorder) {
117 ca->packetcount = 0;
118 ca->lasttime = now;
119 return;
122 ca->packetcount += pkts_acked;
124 if (ca->packetcount >= tp->snd_cwnd - (ca->alpha >> 7 ? : 1)
125 && now - ca->lasttime >= ca->minRTT
126 && ca->minRTT > 0) {
127 __u32 cur_Bi = ca->packetcount * HZ / (now - ca->lasttime);
129 if (htcp_ccount(ca) <= 3) {
130 /* just after backoff */
131 ca->minB = ca->maxB = ca->Bi = cur_Bi;
132 } else {
133 ca->Bi = (3 * ca->Bi + cur_Bi) / 4;
134 if (ca->Bi > ca->maxB)
135 ca->maxB = ca->Bi;
136 if (ca->minB > ca->maxB)
137 ca->minB = ca->maxB;
139 ca->packetcount = 0;
140 ca->lasttime = now;
144 static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
146 if (use_bandwidth_switch) {
147 u32 maxB = ca->maxB;
148 u32 old_maxB = ca->old_maxB;
149 ca->old_maxB = ca->maxB;
151 if (!between(5 * maxB, 4 * old_maxB, 6 * old_maxB)) {
152 ca->beta = BETA_MIN;
153 ca->modeswitch = 0;
154 return;
158 if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
159 ca->beta = (minRTT << 7) / maxRTT;
160 if (ca->beta < BETA_MIN)
161 ca->beta = BETA_MIN;
162 else if (ca->beta > BETA_MAX)
163 ca->beta = BETA_MAX;
164 } else {
165 ca->beta = BETA_MIN;
166 ca->modeswitch = 1;
170 static inline void htcp_alpha_update(struct htcp *ca)
172 u32 minRTT = ca->minRTT;
173 u32 factor = 1;
174 u32 diff = htcp_cong_time(ca);
176 if (diff > HZ) {
177 diff -= HZ;
178 factor = 1 + (10 * diff + ((diff / 2) * (diff / 2) / HZ)) / HZ;
181 if (use_rtt_scaling && minRTT) {
182 u32 scale = (HZ << 3) / (10 * minRTT);
184 /* clamping ratio to interval [0.5,10]<<3 */
185 scale = min(max(scale, 1U << 2), 10U << 3);
186 factor = (factor << 3) / scale;
187 if (!factor)
188 factor = 1;
191 ca->alpha = 2 * factor * ((1 << 7) - ca->beta);
192 if (!ca->alpha)
193 ca->alpha = ALPHA_BASE;
197 * After we have the rtt data to calculate beta, we'd still prefer to wait one
198 * rtt before we adjust our beta to ensure we are working from a consistent
199 * data.
201 * This function should be called when we hit a congestion event since only at
202 * that point do we really have a real sense of maxRTT (the queues en route
203 * were getting just too full now).
205 static void htcp_param_update(struct sock *sk)
207 struct htcp *ca = inet_csk_ca(sk);
208 u32 minRTT = ca->minRTT;
209 u32 maxRTT = ca->maxRTT;
211 htcp_beta_update(ca, minRTT, maxRTT);
212 htcp_alpha_update(ca);
214 /* add slowly fading memory for maxRTT to accommodate routing changes */
215 if (minRTT > 0 && maxRTT > minRTT)
216 ca->maxRTT = minRTT + ((maxRTT - minRTT) * 95) / 100;
219 static u32 htcp_recalc_ssthresh(struct sock *sk)
221 const struct tcp_sock *tp = tcp_sk(sk);
222 const struct htcp *ca = inet_csk_ca(sk);
224 htcp_param_update(sk);
225 return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
228 static void htcp_cong_avoid(struct sock *sk, u32 ack,
229 u32 in_flight, int data_acked)
231 struct tcp_sock *tp = tcp_sk(sk);
232 struct htcp *ca = inet_csk_ca(sk);
234 if (!tcp_is_cwnd_limited(sk, in_flight))
235 return;
237 if (tp->snd_cwnd <= tp->snd_ssthresh)
238 tcp_slow_start(tp);
239 else {
240 /* In dangerous area, increase slowly.
241 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
243 if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
244 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
245 tp->snd_cwnd++;
246 tp->snd_cwnd_cnt = 0;
247 htcp_alpha_update(ca);
248 } else
249 tp->snd_cwnd_cnt += ca->pkts_acked;
251 ca->pkts_acked = 1;
255 static void htcp_init(struct sock *sk)
257 struct htcp *ca = inet_csk_ca(sk);
259 memset(ca, 0, sizeof(struct htcp));
260 ca->alpha = ALPHA_BASE;
261 ca->beta = BETA_MIN;
262 ca->pkts_acked = 1;
263 ca->last_cong = jiffies;
266 static void htcp_state(struct sock *sk, u8 new_state)
268 switch (new_state) {
269 case TCP_CA_Open:
271 struct htcp *ca = inet_csk_ca(sk);
272 ca->last_cong = jiffies;
274 break;
275 case TCP_CA_CWR:
276 case TCP_CA_Recovery:
277 case TCP_CA_Loss:
278 htcp_reset(inet_csk_ca(sk));
279 break;
283 static struct tcp_congestion_ops htcp = {
284 .init = htcp_init,
285 .ssthresh = htcp_recalc_ssthresh,
286 .cong_avoid = htcp_cong_avoid,
287 .set_state = htcp_state,
288 .undo_cwnd = htcp_cwnd_undo,
289 .pkts_acked = measure_achieved_throughput,
290 .owner = THIS_MODULE,
291 .name = "htcp",
294 static int __init htcp_register(void)
296 BUILD_BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
297 BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
298 return tcp_register_congestion_control(&htcp);
301 static void __exit htcp_unregister(void)
303 tcp_unregister_congestion_control(&htcp);
306 module_init(htcp_register);
307 module_exit(htcp_unregister);
309 MODULE_AUTHOR("Baruch Even");
310 MODULE_LICENSE("GPL");
311 MODULE_DESCRIPTION("H-TCP");