blob: 1020eb48d8d18590a0ae4da87fc8462bdb7fd60f [file] [log] [blame]
Baruch Evena7868ea2005-06-23 12:28:11 -07001/*
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 */
8
Baruch Evena7868ea2005-06-23 12:28:11 -07009#include <linux/mm.h>
10#include <linux/module.h>
11#include <net/tcp.h>
12
Stephen Hemminger9121c7772007-02-12 13:34:03 -080013#define ALPHA_BASE (1<<7) /* 1.0 with shift << 7 */
14#define BETA_MIN (1<<6) /* 0.5 with shift << 7 */
Baruch Evena7868ea2005-06-23 12:28:11 -070015#define BETA_MAX 102 /* 0.8 with shift << 7 */
16
Stephen Hemminger9121c7772007-02-12 13:34:03 -080017static int use_rtt_scaling __read_mostly = 1;
Baruch Evena7868ea2005-06-23 12:28:11 -070018module_param(use_rtt_scaling, int, 0644);
19MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling");
20
Stephen Hemminger9121c7772007-02-12 13:34:03 -080021static int use_bandwidth_switch __read_mostly = 1;
Baruch Evena7868ea2005-06-23 12:28:11 -070022module_param(use_bandwidth_switch, int, 0644);
23MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher");
24
25struct htcp {
Gavin McCullagh2a272f92006-10-25 23:05:52 -070026 u32 alpha; /* Fixed point arith, << 7 */
Baruch Evena7868ea2005-06-23 12:28:11 -070027 u8 beta; /* Fixed point arith, << 7 */
Stephen Hemminger9121c7772007-02-12 13:34:03 -080028 u8 modeswitch; /* Delay modeswitch
29 until we had at least one congestion event */
Baruch Even0bc6d902006-03-20 22:22:47 -080030 u16 pkts_acked;
31 u32 packetcount;
Baruch Evena7868ea2005-06-23 12:28:11 -070032 u32 minRTT;
33 u32 maxRTT;
David S. Miller24040432006-11-10 15:01:14 -080034 u32 last_cong; /* Time since last congestion event end */
35 u32 undo_last_cong;
Baruch Evena7868ea2005-06-23 12:28:11 -070036
37 u32 undo_maxRTT;
38 u32 undo_old_maxB;
39
40 /* Bandwidth estimation */
41 u32 minB;
42 u32 maxB;
43 u32 old_maxB;
44 u32 Bi;
45 u32 lasttime;
46};
47
Stephen Hemminger9121c7772007-02-12 13:34:03 -080048static inline u32 htcp_cong_time(const struct htcp *ca)
Baruch Even50bf3e22006-03-20 22:23:10 -080049{
50 return jiffies - ca->last_cong;
51}
52
Stephen Hemminger9121c7772007-02-12 13:34:03 -080053static inline u32 htcp_ccount(const struct htcp *ca)
Baruch Even50bf3e22006-03-20 22:23:10 -080054{
Stephen Hemminger9121c7772007-02-12 13:34:03 -080055 return htcp_cong_time(ca) / ca->minRTT;
Baruch Even50bf3e22006-03-20 22:23:10 -080056}
57
Baruch Evena7868ea2005-06-23 12:28:11 -070058static inline void htcp_reset(struct htcp *ca)
59{
Baruch Even50bf3e22006-03-20 22:23:10 -080060 ca->undo_last_cong = ca->last_cong;
Baruch Evena7868ea2005-06-23 12:28:11 -070061 ca->undo_maxRTT = ca->maxRTT;
62 ca->undo_old_maxB = ca->old_maxB;
63
Baruch Even50bf3e22006-03-20 22:23:10 -080064 ca->last_cong = jiffies;
Baruch Evena7868ea2005-06-23 12:28:11 -070065}
66
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030067static u32 htcp_cwnd_undo(struct sock *sk)
Baruch Evena7868ea2005-06-23 12:28:11 -070068{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030069 const struct tcp_sock *tp = tcp_sk(sk);
70 struct htcp *ca = inet_csk_ca(sk);
Stephen Hemminger9121c7772007-02-12 13:34:03 -080071
Baruch Even50bf3e22006-03-20 22:23:10 -080072 ca->last_cong = ca->undo_last_cong;
Baruch Evena7868ea2005-06-23 12:28:11 -070073 ca->maxRTT = ca->undo_maxRTT;
74 ca->old_maxB = ca->undo_old_maxB;
Stephen Hemminger9121c7772007-02-12 13:34:03 -080075
76 return max(tp->snd_cwnd, (tp->snd_ssthresh << 7) / ca->beta);
Baruch Evena7868ea2005-06-23 12:28:11 -070077}
78
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030079static inline void measure_rtt(struct sock *sk)
Baruch Evena7868ea2005-06-23 12:28:11 -070080{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030081 const struct inet_connection_sock *icsk = inet_csk(sk);
82 const struct tcp_sock *tp = tcp_sk(sk);
83 struct htcp *ca = inet_csk_ca(sk);
Stephen Hemminger9121c7772007-02-12 13:34:03 -080084 u32 srtt = tp->srtt >> 3;
Baruch Evena7868ea2005-06-23 12:28:11 -070085
86 /* keep track of minimum RTT seen so far, minRTT is zero at first */
87 if (ca->minRTT > srtt || !ca->minRTT)
88 ca->minRTT = srtt;
89
90 /* max RTT */
Stephen Hemminger9121c7772007-02-12 13:34:03 -080091 if (icsk->icsk_ca_state == TCP_CA_Open
92 && tp->snd_ssthresh < 0xFFFF && htcp_ccount(ca) > 3) {
Baruch Evena7868ea2005-06-23 12:28:11 -070093 if (ca->maxRTT < ca->minRTT)
94 ca->maxRTT = ca->minRTT;
Stephen Hemminger9121c7772007-02-12 13:34:03 -080095 if (ca->maxRTT < srtt
96 && srtt <= ca->maxRTT + msecs_to_jiffies(20))
Baruch Evena7868ea2005-06-23 12:28:11 -070097 ca->maxRTT = srtt;
98 }
99}
100
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300101static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked)
Baruch Evena7868ea2005-06-23 12:28:11 -0700102{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300103 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);
Baruch Evena7868ea2005-06-23 12:28:11 -0700106 u32 now = tcp_time_stamp;
107
Baruch Even0bc6d902006-03-20 22:22:47 -0800108 if (icsk->icsk_ca_state == TCP_CA_Open)
109 ca->pkts_acked = pkts_acked;
110
111 if (!use_bandwidth_switch)
112 return;
113
Baruch Evena7868ea2005-06-23 12:28:11 -0700114 /* achieved throughput calculations */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300115 if (icsk->icsk_ca_state != TCP_CA_Open &&
116 icsk->icsk_ca_state != TCP_CA_Disorder) {
Baruch Evena7868ea2005-06-23 12:28:11 -0700117 ca->packetcount = 0;
118 ca->lasttime = now;
119 return;
120 }
121
122 ca->packetcount += pkts_acked;
123
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800124 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);
128
Baruch Even50bf3e22006-03-20 22:23:10 -0800129 if (htcp_ccount(ca) <= 3) {
Baruch Evena7868ea2005-06-23 12:28:11 -0700130 /* just after backoff */
131 ca->minB = ca->maxB = ca->Bi = cur_Bi;
132 } else {
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800133 ca->Bi = (3 * ca->Bi + cur_Bi) / 4;
Baruch Evena7868ea2005-06-23 12:28:11 -0700134 if (ca->Bi > ca->maxB)
135 ca->maxB = ca->Bi;
136 if (ca->minB > ca->maxB)
137 ca->minB = ca->maxB;
138 }
139 ca->packetcount = 0;
140 ca->lasttime = now;
141 }
142}
143
144static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
145{
146 if (use_bandwidth_switch) {
147 u32 maxB = ca->maxB;
148 u32 old_maxB = ca->old_maxB;
149 ca->old_maxB = ca->maxB;
150
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800151 if (!between(5 * maxB, 4 * old_maxB, 6 * old_maxB)) {
Baruch Evena7868ea2005-06-23 12:28:11 -0700152 ca->beta = BETA_MIN;
153 ca->modeswitch = 0;
154 return;
155 }
156 }
157
Baruch Evenc33ad6e2006-03-20 22:22:20 -0800158 if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800159 ca->beta = (minRTT << 7) / maxRTT;
Baruch Evena7868ea2005-06-23 12:28:11 -0700160 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;
167 }
168}
169
170static inline void htcp_alpha_update(struct htcp *ca)
171{
172 u32 minRTT = ca->minRTT;
173 u32 factor = 1;
Baruch Even50bf3e22006-03-20 22:23:10 -0800174 u32 diff = htcp_cong_time(ca);
Baruch Evena7868ea2005-06-23 12:28:11 -0700175
176 if (diff > HZ) {
177 diff -= HZ;
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800178 factor = 1 + (10 * diff + ((diff / 2) * (diff / 2) / HZ)) / HZ;
Baruch Evena7868ea2005-06-23 12:28:11 -0700179 }
180
181 if (use_rtt_scaling && minRTT) {
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800182 u32 scale = (HZ << 3) / (10 * minRTT);
183
184 /* clamping ratio to interval [0.5,10]<<3 */
185 scale = min(max(scale, 1U << 2), 10U << 3);
186 factor = (factor << 3) / scale;
Baruch Evena7868ea2005-06-23 12:28:11 -0700187 if (!factor)
188 factor = 1;
189 }
190
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800191 ca->alpha = 2 * factor * ((1 << 7) - ca->beta);
Baruch Evena7868ea2005-06-23 12:28:11 -0700192 if (!ca->alpha)
193 ca->alpha = ALPHA_BASE;
194}
195
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800196/*
197 * After we have the rtt data to calculate beta, we'd still prefer to wait one
Baruch Evena7868ea2005-06-23 12:28:11 -0700198 * rtt before we adjust our beta to ensure we are working from a consistent
199 * data.
200 *
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).
204 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300205static void htcp_param_update(struct sock *sk)
Baruch Evena7868ea2005-06-23 12:28:11 -0700206{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300207 struct htcp *ca = inet_csk_ca(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -0700208 u32 minRTT = ca->minRTT;
209 u32 maxRTT = ca->maxRTT;
210
211 htcp_beta_update(ca, minRTT, maxRTT);
212 htcp_alpha_update(ca);
213
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800214 /* add slowly fading memory for maxRTT to accommodate routing changes */
Baruch Evena7868ea2005-06-23 12:28:11 -0700215 if (minRTT > 0 && maxRTT > minRTT)
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800216 ca->maxRTT = minRTT + ((maxRTT - minRTT) * 95) / 100;
Baruch Evena7868ea2005-06-23 12:28:11 -0700217}
218
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300219static u32 htcp_recalc_ssthresh(struct sock *sk)
Baruch Evena7868ea2005-06-23 12:28:11 -0700220{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300221 const struct tcp_sock *tp = tcp_sk(sk);
222 const struct htcp *ca = inet_csk_ca(sk);
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800223
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300224 htcp_param_update(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -0700225 return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
226}
227
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300228static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
Baruch Evena7868ea2005-06-23 12:28:11 -0700229 u32 in_flight, int data_acked)
230{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300231 struct tcp_sock *tp = tcp_sk(sk);
232 struct htcp *ca = inet_csk_ca(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -0700233
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -0800234 if (!tcp_is_cwnd_limited(sk, in_flight))
Baruch Evena7868ea2005-06-23 12:28:11 -0700235 return;
236
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900237 if (tp->snd_cwnd <= tp->snd_ssthresh)
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800238 tcp_slow_start(tp);
239 else {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300240 measure_rtt(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -0700241
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800242 /* In dangerous area, increase slowly.
Baruch Evena7868ea2005-06-23 12:28:11 -0700243 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
244 */
Baruch Even0bc6d902006-03-20 22:22:47 -0800245 if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
Baruch Evena7868ea2005-06-23 12:28:11 -0700246 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
247 tp->snd_cwnd++;
248 tp->snd_cwnd_cnt = 0;
Baruch Even50bf3e22006-03-20 22:23:10 -0800249 htcp_alpha_update(ca);
Baruch Even0bc6d902006-03-20 22:22:47 -0800250 } else
251 tp->snd_cwnd_cnt += ca->pkts_acked;
252
253 ca->pkts_acked = 1;
Baruch Evena7868ea2005-06-23 12:28:11 -0700254 }
255}
256
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300257static void htcp_init(struct sock *sk)
Baruch Evena7868ea2005-06-23 12:28:11 -0700258{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300259 struct htcp *ca = inet_csk_ca(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -0700260
261 memset(ca, 0, sizeof(struct htcp));
262 ca->alpha = ALPHA_BASE;
263 ca->beta = BETA_MIN;
Baruch Even0bc6d902006-03-20 22:22:47 -0800264 ca->pkts_acked = 1;
Baruch Even50bf3e22006-03-20 22:23:10 -0800265 ca->last_cong = jiffies;
Baruch Evena7868ea2005-06-23 12:28:11 -0700266}
267
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300268static void htcp_state(struct sock *sk, u8 new_state)
Baruch Evena7868ea2005-06-23 12:28:11 -0700269{
270 switch (new_state) {
Baruch Even50bf3e22006-03-20 22:23:10 -0800271 case TCP_CA_Open:
272 {
273 struct htcp *ca = inet_csk_ca(sk);
274 ca->last_cong = jiffies;
275 }
276 break;
Baruch Evena7868ea2005-06-23 12:28:11 -0700277 case TCP_CA_CWR:
278 case TCP_CA_Recovery:
279 case TCP_CA_Loss:
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300280 htcp_reset(inet_csk_ca(sk));
Baruch Evena7868ea2005-06-23 12:28:11 -0700281 break;
282 }
283}
284
285static struct tcp_congestion_ops htcp = {
286 .init = htcp_init,
287 .ssthresh = htcp_recalc_ssthresh,
Baruch Evena7868ea2005-06-23 12:28:11 -0700288 .cong_avoid = htcp_cong_avoid,
289 .set_state = htcp_state,
290 .undo_cwnd = htcp_cwnd_undo,
291 .pkts_acked = measure_achieved_throughput,
292 .owner = THIS_MODULE,
293 .name = "htcp",
294};
295
296static int __init htcp_register(void)
297{
Alexey Dobriyan74975d42006-08-25 17:10:33 -0700298 BUILD_BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
Baruch Evena7868ea2005-06-23 12:28:11 -0700299 BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
Baruch Evena7868ea2005-06-23 12:28:11 -0700300 return tcp_register_congestion_control(&htcp);
301}
302
303static void __exit htcp_unregister(void)
304{
305 tcp_unregister_congestion_control(&htcp);
306}
307
308module_init(htcp_register);
309module_exit(htcp_unregister);
310
311MODULE_AUTHOR("Baruch Even");
312MODULE_LICENSE("GPL");
313MODULE_DESCRIPTION("H-TCP");