blob: 7c94a49554167a7315c22faf7cbaafa13824c434 [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
Doug Leith8f65b532008-11-12 01:41:09 -080072 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;
77 }
Stephen Hemminger9121c7772007-02-12 13:34:03 -080078
79 return max(tp->snd_cwnd, (tp->snd_ssthresh << 7) / ca->beta);
Baruch Evena7868ea2005-06-23 12:28:11 -070080}
81
Stephen Hemminger113bbbd2007-07-25 23:50:28 -070082static inline void measure_rtt(struct sock *sk, u32 srtt)
Baruch Evena7868ea2005-06-23 12:28:11 -070083{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030084 const struct inet_connection_sock *icsk = inet_csk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030085 struct htcp *ca = inet_csk_ca(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -070086
87 /* keep track of minimum RTT seen so far, minRTT is zero at first */
88 if (ca->minRTT > srtt || !ca->minRTT)
89 ca->minRTT = srtt;
90
91 /* max RTT */
Stephen Hemmingerf34d1952007-08-07 18:29:05 -070092 if (icsk->icsk_ca_state == TCP_CA_Open) {
Baruch Evena7868ea2005-06-23 12:28:11 -070093 if (ca->maxRTT < ca->minRTT)
94 ca->maxRTT = ca->minRTT;
Joe Perches9d4fb272009-11-23 10:41:23 -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
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -0700101static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked, s32 rtt)
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
Stephen Hemminger113bbbd2007-07-25 23:50:28 -0700111 if (rtt > 0)
112 measure_rtt(sk, usecs_to_jiffies(rtt));
113
Baruch Even0bc6d902006-03-20 22:22:47 -0800114 if (!use_bandwidth_switch)
115 return;
116
Baruch Evena7868ea2005-06-23 12:28:11 -0700117 /* achieved throughput calculations */
Ilpo Järvinen571a5dd2009-02-28 04:44:36 +0000118 if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_Disorder))) {
Baruch Evena7868ea2005-06-23 12:28:11 -0700119 ca->packetcount = 0;
120 ca->lasttime = now;
121 return;
122 }
123
124 ca->packetcount += pkts_acked;
125
Joe Perches9d4fb272009-11-23 10:41:23 -0800126 if (ca->packetcount >= tp->snd_cwnd - (ca->alpha >> 7 ? : 1) &&
127 now - ca->lasttime >= ca->minRTT &&
128 ca->minRTT > 0) {
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800129 __u32 cur_Bi = ca->packetcount * HZ / (now - ca->lasttime);
130
Baruch Even50bf3e22006-03-20 22:23:10 -0800131 if (htcp_ccount(ca) <= 3) {
Baruch Evena7868ea2005-06-23 12:28:11 -0700132 /* just after backoff */
133 ca->minB = ca->maxB = ca->Bi = cur_Bi;
134 } else {
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800135 ca->Bi = (3 * ca->Bi + cur_Bi) / 4;
Baruch Evena7868ea2005-06-23 12:28:11 -0700136 if (ca->Bi > ca->maxB)
137 ca->maxB = ca->Bi;
138 if (ca->minB > ca->maxB)
139 ca->minB = ca->maxB;
140 }
141 ca->packetcount = 0;
142 ca->lasttime = now;
143 }
144}
145
146static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
147{
148 if (use_bandwidth_switch) {
149 u32 maxB = ca->maxB;
150 u32 old_maxB = ca->old_maxB;
151 ca->old_maxB = ca->maxB;
152
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800153 if (!between(5 * maxB, 4 * old_maxB, 6 * old_maxB)) {
Baruch Evena7868ea2005-06-23 12:28:11 -0700154 ca->beta = BETA_MIN;
155 ca->modeswitch = 0;
156 return;
157 }
158 }
159
Baruch Evenc33ad6e2006-03-20 22:22:20 -0800160 if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800161 ca->beta = (minRTT << 7) / maxRTT;
Baruch Evena7868ea2005-06-23 12:28:11 -0700162 if (ca->beta < BETA_MIN)
163 ca->beta = BETA_MIN;
164 else if (ca->beta > BETA_MAX)
165 ca->beta = BETA_MAX;
166 } else {
167 ca->beta = BETA_MIN;
168 ca->modeswitch = 1;
169 }
170}
171
172static inline void htcp_alpha_update(struct htcp *ca)
173{
174 u32 minRTT = ca->minRTT;
175 u32 factor = 1;
Baruch Even50bf3e22006-03-20 22:23:10 -0800176 u32 diff = htcp_cong_time(ca);
Baruch Evena7868ea2005-06-23 12:28:11 -0700177
178 if (diff > HZ) {
179 diff -= HZ;
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800180 factor = 1 + (10 * diff + ((diff / 2) * (diff / 2) / HZ)) / HZ;
Baruch Evena7868ea2005-06-23 12:28:11 -0700181 }
182
183 if (use_rtt_scaling && minRTT) {
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800184 u32 scale = (HZ << 3) / (10 * minRTT);
185
186 /* clamping ratio to interval [0.5,10]<<3 */
187 scale = min(max(scale, 1U << 2), 10U << 3);
188 factor = (factor << 3) / scale;
Baruch Evena7868ea2005-06-23 12:28:11 -0700189 if (!factor)
190 factor = 1;
191 }
192
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800193 ca->alpha = 2 * factor * ((1 << 7) - ca->beta);
Baruch Evena7868ea2005-06-23 12:28:11 -0700194 if (!ca->alpha)
195 ca->alpha = ALPHA_BASE;
196}
197
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800198/*
199 * After we have the rtt data to calculate beta, we'd still prefer to wait one
Baruch Evena7868ea2005-06-23 12:28:11 -0700200 * rtt before we adjust our beta to ensure we are working from a consistent
201 * data.
202 *
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).
206 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300207static void htcp_param_update(struct sock *sk)
Baruch Evena7868ea2005-06-23 12:28:11 -0700208{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300209 struct htcp *ca = inet_csk_ca(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -0700210 u32 minRTT = ca->minRTT;
211 u32 maxRTT = ca->maxRTT;
212
213 htcp_beta_update(ca, minRTT, maxRTT);
214 htcp_alpha_update(ca);
215
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800216 /* add slowly fading memory for maxRTT to accommodate routing changes */
Baruch Evena7868ea2005-06-23 12:28:11 -0700217 if (minRTT > 0 && maxRTT > minRTT)
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800218 ca->maxRTT = minRTT + ((maxRTT - minRTT) * 95) / 100;
Baruch Evena7868ea2005-06-23 12:28:11 -0700219}
220
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300221static u32 htcp_recalc_ssthresh(struct sock *sk)
Baruch Evena7868ea2005-06-23 12:28:11 -0700222{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300223 const struct tcp_sock *tp = tcp_sk(sk);
224 const struct htcp *ca = inet_csk_ca(sk);
Stephen Hemminger9121c7772007-02-12 13:34:03 -0800225
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300226 htcp_param_update(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -0700227 return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
228}
229
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +0200230static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Baruch Evena7868ea2005-06-23 12:28:11 -0700231{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300232 struct tcp_sock *tp = tcp_sk(sk);
233 struct htcp *ca = inet_csk_ca(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -0700234
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -0800235 if (!tcp_is_cwnd_limited(sk, in_flight))
Baruch Evena7868ea2005-06-23 12:28:11 -0700236 return;
237
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900238 if (tp->snd_cwnd <= tp->snd_ssthresh)
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800239 tcp_slow_start(tp);
240 else {
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800241 /* In dangerous area, increase slowly.
Baruch Evena7868ea2005-06-23 12:28:11 -0700242 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
243 */
Baruch Even0bc6d902006-03-20 22:22:47 -0800244 if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
Baruch Evena7868ea2005-06-23 12:28:11 -0700245 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
246 tp->snd_cwnd++;
247 tp->snd_cwnd_cnt = 0;
Baruch Even50bf3e22006-03-20 22:23:10 -0800248 htcp_alpha_update(ca);
Baruch Even0bc6d902006-03-20 22:22:47 -0800249 } else
250 tp->snd_cwnd_cnt += ca->pkts_acked;
251
252 ca->pkts_acked = 1;
Baruch Evena7868ea2005-06-23 12:28:11 -0700253 }
254}
255
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300256static void htcp_init(struct sock *sk)
Baruch Evena7868ea2005-06-23 12:28:11 -0700257{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300258 struct htcp *ca = inet_csk_ca(sk);
Baruch Evena7868ea2005-06-23 12:28:11 -0700259
260 memset(ca, 0, sizeof(struct htcp));
261 ca->alpha = ALPHA_BASE;
262 ca->beta = BETA_MIN;
Baruch Even0bc6d902006-03-20 22:22:47 -0800263 ca->pkts_acked = 1;
Baruch Even50bf3e22006-03-20 22:23:10 -0800264 ca->last_cong = jiffies;
Baruch Evena7868ea2005-06-23 12:28:11 -0700265}
266
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300267static void htcp_state(struct sock *sk, u8 new_state)
Baruch Evena7868ea2005-06-23 12:28:11 -0700268{
269 switch (new_state) {
Baruch Even50bf3e22006-03-20 22:23:10 -0800270 case TCP_CA_Open:
271 {
272 struct htcp *ca = inet_csk_ca(sk);
Doug Leith8f65b532008-11-12 01:41:09 -0800273 if (ca->undo_last_cong) {
274 ca->last_cong = jiffies;
275 ca->undo_last_cong = 0;
276 }
Baruch Even50bf3e22006-03-20 22:23:10 -0800277 }
278 break;
Baruch Evena7868ea2005-06-23 12:28:11 -0700279 case TCP_CA_CWR:
280 case TCP_CA_Recovery:
281 case TCP_CA_Loss:
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300282 htcp_reset(inet_csk_ca(sk));
Baruch Evena7868ea2005-06-23 12:28:11 -0700283 break;
284 }
285}
286
287static struct tcp_congestion_ops htcp = {
288 .init = htcp_init,
289 .ssthresh = htcp_recalc_ssthresh,
Baruch Evena7868ea2005-06-23 12:28:11 -0700290 .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,
295 .name = "htcp",
296};
297
298static int __init htcp_register(void)
299{
Alexey Dobriyan74975d42006-08-25 17:10:33 -0700300 BUILD_BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
Baruch Evena7868ea2005-06-23 12:28:11 -0700301 BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
Baruch Evena7868ea2005-06-23 12:28:11 -0700302 return tcp_register_congestion_control(&htcp);
303}
304
305static void __exit htcp_unregister(void)
306{
307 tcp_unregister_congestion_control(&htcp);
308}
309
310module_init(htcp_register);
311module_exit(htcp_unregister);
312
313MODULE_AUTHOR("Baruch Even");
314MODULE_LICENSE("GPL");
315MODULE_DESCRIPTION("H-TCP");