[PATCH] dm flush queue EINTR
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / net / ipv4 / tcp_htcp.c
blob1b2ff53f98ed2cb2801429c707579fac66baab5a
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/config.h>
10 #include <linux/mm.h>
11 #include <linux/module.h>
12 #include <net/tcp.h>
14 #define ALPHA_BASE (1<<7) /* 1.0 with shift << 7 */
15 #define BETA_MIN (1<<6) /* 0.5 with shift << 7 */
16 #define BETA_MAX 102 /* 0.8 with shift << 7 */
18 static int use_rtt_scaling = 1;
19 module_param(use_rtt_scaling, int, 0644);
20 MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling");
22 static int use_bandwidth_switch = 1;
23 module_param(use_bandwidth_switch, int, 0644);
24 MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher");
26 struct htcp {
27 u16 alpha; /* Fixed point arith, << 7 */
28 u8 beta; /* Fixed point arith, << 7 */
29 u8 modeswitch; /* Delay modeswitch until we had at least one congestion event */
30 u32 last_cong; /* Time since last congestion event end */
31 u32 undo_last_cong;
32 u16 pkts_acked;
33 u32 packetcount;
34 u32 minRTT;
35 u32 maxRTT;
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(struct htcp *ca)
50 return jiffies - ca->last_cong;
53 static inline u32 htcp_ccount(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);
71 ca->last_cong = ca->undo_last_cong;
72 ca->maxRTT = ca->undo_maxRTT;
73 ca->old_maxB = ca->undo_old_maxB;
74 return max(tp->snd_cwnd, (tp->snd_ssthresh<<7)/ca->beta);
77 static inline void measure_rtt(struct sock *sk)
79 const struct inet_connection_sock *icsk = inet_csk(sk);
80 const struct tcp_sock *tp = tcp_sk(sk);
81 struct htcp *ca = inet_csk_ca(sk);
82 u32 srtt = tp->srtt>>3;
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 && tp->snd_ssthresh < 0xFFFF && htcp_ccount(ca) > 3) {
90 if (ca->maxRTT < ca->minRTT)
91 ca->maxRTT = ca->minRTT;
92 if (ca->maxRTT < srtt && srtt <= ca->maxRTT+msecs_to_jiffies(20))
93 ca->maxRTT = srtt;
97 static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked)
99 const struct inet_connection_sock *icsk = inet_csk(sk);
100 const struct tcp_sock *tp = tcp_sk(sk);
101 struct htcp *ca = inet_csk_ca(sk);
102 u32 now = tcp_time_stamp;
104 if (icsk->icsk_ca_state == TCP_CA_Open)
105 ca->pkts_acked = pkts_acked;
107 if (!use_bandwidth_switch)
108 return;
110 /* achieved throughput calculations */
111 if (icsk->icsk_ca_state != TCP_CA_Open &&
112 icsk->icsk_ca_state != TCP_CA_Disorder) {
113 ca->packetcount = 0;
114 ca->lasttime = now;
115 return;
118 ca->packetcount += pkts_acked;
120 if (ca->packetcount >= tp->snd_cwnd - (ca->alpha>>7? : 1)
121 && now - ca->lasttime >= ca->minRTT
122 && ca->minRTT > 0) {
123 __u32 cur_Bi = ca->packetcount*HZ/(now - ca->lasttime);
124 if (htcp_ccount(ca) <= 3) {
125 /* just after backoff */
126 ca->minB = ca->maxB = ca->Bi = cur_Bi;
127 } else {
128 ca->Bi = (3*ca->Bi + cur_Bi)/4;
129 if (ca->Bi > ca->maxB)
130 ca->maxB = ca->Bi;
131 if (ca->minB > ca->maxB)
132 ca->minB = ca->maxB;
134 ca->packetcount = 0;
135 ca->lasttime = now;
139 static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
141 if (use_bandwidth_switch) {
142 u32 maxB = ca->maxB;
143 u32 old_maxB = ca->old_maxB;
144 ca->old_maxB = ca->maxB;
146 if (!between(5*maxB, 4*old_maxB, 6*old_maxB)) {
147 ca->beta = BETA_MIN;
148 ca->modeswitch = 0;
149 return;
153 if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
154 ca->beta = (minRTT<<7)/maxRTT;
155 if (ca->beta < BETA_MIN)
156 ca->beta = BETA_MIN;
157 else if (ca->beta > BETA_MAX)
158 ca->beta = BETA_MAX;
159 } else {
160 ca->beta = BETA_MIN;
161 ca->modeswitch = 1;
165 static inline void htcp_alpha_update(struct htcp *ca)
167 u32 minRTT = ca->minRTT;
168 u32 factor = 1;
169 u32 diff = htcp_cong_time(ca);
171 if (diff > HZ) {
172 diff -= HZ;
173 factor = 1+ ( 10*diff + ((diff/2)*(diff/2)/HZ) )/HZ;
176 if (use_rtt_scaling && minRTT) {
177 u32 scale = (HZ<<3)/(10*minRTT);
178 scale = min(max(scale, 1U<<2), 10U<<3); /* clamping ratio to interval [0.5,10]<<3 */
179 factor = (factor<<3)/scale;
180 if (!factor)
181 factor = 1;
184 ca->alpha = 2*factor*((1<<7)-ca->beta);
185 if (!ca->alpha)
186 ca->alpha = ALPHA_BASE;
189 /* After we have the rtt data to calculate beta, we'd still prefer to wait one
190 * rtt before we adjust our beta to ensure we are working from a consistent
191 * data.
193 * This function should be called when we hit a congestion event since only at
194 * that point do we really have a real sense of maxRTT (the queues en route
195 * were getting just too full now).
197 static void htcp_param_update(struct sock *sk)
199 struct htcp *ca = inet_csk_ca(sk);
200 u32 minRTT = ca->minRTT;
201 u32 maxRTT = ca->maxRTT;
203 htcp_beta_update(ca, minRTT, maxRTT);
204 htcp_alpha_update(ca);
206 /* add slowly fading memory for maxRTT to accommodate routing changes etc */
207 if (minRTT > 0 && maxRTT > minRTT)
208 ca->maxRTT = minRTT + ((maxRTT-minRTT)*95)/100;
211 static u32 htcp_recalc_ssthresh(struct sock *sk)
213 const struct tcp_sock *tp = tcp_sk(sk);
214 const struct htcp *ca = inet_csk_ca(sk);
215 htcp_param_update(sk);
216 return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
219 static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
220 u32 in_flight, int data_acked)
222 struct tcp_sock *tp = tcp_sk(sk);
223 struct htcp *ca = inet_csk_ca(sk);
225 if (!tcp_is_cwnd_limited(sk, in_flight))
226 return;
228 if (tp->snd_cwnd <= tp->snd_ssthresh)
229 tcp_slow_start(tp);
230 else {
232 measure_rtt(sk);
234 /* In dangerous area, increase slowly.
235 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
237 if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
238 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
239 tp->snd_cwnd++;
240 tp->snd_cwnd_cnt = 0;
241 htcp_alpha_update(ca);
242 } else
243 tp->snd_cwnd_cnt += ca->pkts_acked;
245 ca->pkts_acked = 1;
249 /* Lower bound on congestion window. */
250 static u32 htcp_min_cwnd(struct sock *sk)
252 const struct tcp_sock *tp = tcp_sk(sk);
253 return tp->snd_ssthresh;
257 static void htcp_init(struct sock *sk)
259 struct htcp *ca = inet_csk_ca(sk);
261 memset(ca, 0, sizeof(struct htcp));
262 ca->alpha = ALPHA_BASE;
263 ca->beta = BETA_MIN;
264 ca->pkts_acked = 1;
265 ca->last_cong = jiffies;
268 static void htcp_state(struct sock *sk, u8 new_state)
270 switch (new_state) {
271 case TCP_CA_Open:
273 struct htcp *ca = inet_csk_ca(sk);
274 ca->last_cong = jiffies;
276 break;
277 case TCP_CA_CWR:
278 case TCP_CA_Recovery:
279 case TCP_CA_Loss:
280 htcp_reset(inet_csk_ca(sk));
281 break;
285 static struct tcp_congestion_ops htcp = {
286 .init = htcp_init,
287 .ssthresh = htcp_recalc_ssthresh,
288 .min_cwnd = htcp_min_cwnd,
289 .cong_avoid = htcp_cong_avoid,
290 .set_state = htcp_state,
291 .undo_cwnd = htcp_cwnd_undo,
292 .pkts_acked = measure_achieved_throughput,
293 .owner = THIS_MODULE,
294 .name = "htcp",
297 static int __init htcp_register(void)
299 BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
300 BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
301 return tcp_register_congestion_control(&htcp);
304 static void __exit htcp_unregister(void)
306 tcp_unregister_congestion_control(&htcp);
309 module_init(htcp_register);
310 module_exit(htcp_unregister);
312 MODULE_AUTHOR("Baruch Even");
313 MODULE_LICENSE("GPL");
314 MODULE_DESCRIPTION("H-TCP");