blob: 4586211e3757f9b30c78b95bdbe1d1a858ea9070 [file] [log] [blame]
Stephen Hemminger83803032005-06-23 12:23:25 -07001/*
2 * Binary Increase Congestion control for TCP
3 *
4 * This is from the implementation of BICTCP in
5 * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
6 * "Binary Increase Congestion Control for Fast, Long Distance
7 * Networks" in InfoComm 2004
8 * Available from:
9 * http://www.csc.ncsu.edu/faculty/rhee/export/bitcp.pdf
10 *
11 * Unless BIC is enabled and congestion window is large
12 * this behaves the same as the original Reno.
13 */
14
Stephen Hemminger83803032005-06-23 12:23:25 -070015#include <linux/mm.h>
16#include <linux/module.h>
17#include <net/tcp.h>
18
19
20#define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
21 * max_cwnd = snd_cwnd * beta
22 */
23#define BICTCP_B 4 /*
24 * In binary search,
25 * go to point (max+min)/N
26 */
27
28static int fast_convergence = 1;
Stephen Hemminger450b5b12005-11-01 15:26:45 -080029static int max_increment = 16;
Stephen Hemminger83803032005-06-23 12:23:25 -070030static int low_window = 14;
31static int beta = 819; /* = 819/1024 (BICTCP_BETA_SCALE) */
David S. Miller66e1e3b2007-06-13 01:03:53 -070032static int initial_ssthresh;
Stephen Hemminger83803032005-06-23 12:23:25 -070033static int smooth_part = 20;
34
35module_param(fast_convergence, int, 0644);
36MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
37module_param(max_increment, int, 0644);
38MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
39module_param(low_window, int, 0644);
40MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
41module_param(beta, int, 0644);
42MODULE_PARM_DESC(beta, "beta for multiplicative increase");
Stephen Hemminger83803032005-06-23 12:23:25 -070043module_param(initial_ssthresh, int, 0644);
44MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
45module_param(smooth_part, int, 0644);
46MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
47
48
49/* BIC TCP Parameters */
50struct bictcp {
51 u32 cnt; /* increase cwnd by 1 after ACKs */
52 u32 last_max_cwnd; /* last maximum snd_cwnd */
53 u32 loss_cwnd; /* congestion window at last loss */
54 u32 last_cwnd; /* the last snd_cwnd */
55 u32 last_time; /* time when updated last_cwnd */
Stephen Hemminger83803032005-06-23 12:23:25 -070056 u32 epoch_start; /* beginning of an epoch */
57#define ACK_RATIO_SHIFT 4
58 u32 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */
59};
60
61static inline void bictcp_reset(struct bictcp *ca)
62{
63 ca->cnt = 0;
64 ca->last_max_cwnd = 0;
65 ca->loss_cwnd = 0;
66 ca->last_cwnd = 0;
67 ca->last_time = 0;
Stephen Hemminger83803032005-06-23 12:23:25 -070068 ca->epoch_start = 0;
69 ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
70}
71
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030072static void bictcp_init(struct sock *sk)
Stephen Hemminger83803032005-06-23 12:23:25 -070073{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030074 bictcp_reset(inet_csk_ca(sk));
Stephen Hemminger83803032005-06-23 12:23:25 -070075 if (initial_ssthresh)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030076 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
Stephen Hemminger83803032005-06-23 12:23:25 -070077}
78
79/*
80 * Compute congestion window to use.
81 */
82static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
83{
84 if (ca->last_cwnd == cwnd &&
85 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
86 return;
87
88 ca->last_cwnd = cwnd;
89 ca->last_time = tcp_time_stamp;
90
91 if (ca->epoch_start == 0) /* record the beginning of an epoch */
92 ca->epoch_start = tcp_time_stamp;
93
94 /* start off normal */
95 if (cwnd <= low_window) {
96 ca->cnt = cwnd;
97 return;
98 }
99
100 /* binary increase */
101 if (cwnd < ca->last_max_cwnd) {
102 __u32 dist = (ca->last_max_cwnd - cwnd)
103 / BICTCP_B;
104
105 if (dist > max_increment)
106 /* linear increase */
107 ca->cnt = cwnd / max_increment;
108 else if (dist <= 1U)
109 /* binary search increase */
110 ca->cnt = (cwnd * smooth_part) / BICTCP_B;
111 else
112 /* binary search increase */
113 ca->cnt = cwnd / dist;
114 } else {
115 /* slow start AMD linear increase */
116 if (cwnd < ca->last_max_cwnd + BICTCP_B)
117 /* slow start */
118 ca->cnt = (cwnd * smooth_part) / BICTCP_B;
119 else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
120 /* slow start */
121 ca->cnt = (cwnd * (BICTCP_B-1))
Stephen Hemminger42a39452005-10-05 12:09:31 -0700122 / (cwnd - ca->last_max_cwnd);
Stephen Hemminger83803032005-06-23 12:23:25 -0700123 else
124 /* linear increase */
125 ca->cnt = cwnd / max_increment;
126 }
127
128 /* if in slow start or link utilization is very low */
Stephen Hemminger018da8f2005-12-13 23:13:00 -0800129 if (ca->loss_cwnd == 0) {
Stephen Hemminger83803032005-06-23 12:23:25 -0700130 if (ca->cnt > 20) /* increase cwnd 5% per RTT */
131 ca->cnt = 20;
132 }
133
134 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
135 if (ca->cnt == 0) /* cannot be zero */
136 ca->cnt = 1;
137}
138
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300139static void bictcp_cong_avoid(struct sock *sk, u32 ack,
Stephen Hemminger16751342007-07-16 18:35:52 -0700140 u32 in_flight, int data_acked)
Stephen Hemminger83803032005-06-23 12:23:25 -0700141{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300142 struct tcp_sock *tp = tcp_sk(sk);
143 struct bictcp *ca = inet_csk_ca(sk);
Stephen Hemminger83803032005-06-23 12:23:25 -0700144
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -0800145 if (!tcp_is_cwnd_limited(sk, in_flight))
Stephen Hemminger83803032005-06-23 12:23:25 -0700146 return;
147
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800148 if (tp->snd_cwnd <= tp->snd_ssthresh)
149 tcp_slow_start(tp);
150 else {
Stephen Hemminger83803032005-06-23 12:23:25 -0700151 bictcp_update(ca, tp->snd_cwnd);
152
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800153 /* In dangerous area, increase slowly.
Stephen Hemminger83803032005-06-23 12:23:25 -0700154 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
155 */
156 if (tp->snd_cwnd_cnt >= ca->cnt) {
157 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
158 tp->snd_cwnd++;
159 tp->snd_cwnd_cnt = 0;
160 } else
161 tp->snd_cwnd_cnt++;
162 }
163
164}
165
166/*
167 * behave like Reno until low_window is reached,
168 * then increase congestion window slowly
169 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300170static u32 bictcp_recalc_ssthresh(struct sock *sk)
Stephen Hemminger83803032005-06-23 12:23:25 -0700171{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300172 const struct tcp_sock *tp = tcp_sk(sk);
173 struct bictcp *ca = inet_csk_ca(sk);
Stephen Hemminger83803032005-06-23 12:23:25 -0700174
175 ca->epoch_start = 0; /* end of epoch */
176
Stephen Hemminger83803032005-06-23 12:23:25 -0700177 /* Wmax and fast convergence */
178 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
179 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
180 / (2 * BICTCP_BETA_SCALE);
181 else
182 ca->last_max_cwnd = tp->snd_cwnd;
183
184 ca->loss_cwnd = tp->snd_cwnd;
185
186
187 if (tp->snd_cwnd <= low_window)
188 return max(tp->snd_cwnd >> 1U, 2U);
189 else
190 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
191}
192
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300193static u32 bictcp_undo_cwnd(struct sock *sk)
Stephen Hemminger83803032005-06-23 12:23:25 -0700194{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300195 const struct tcp_sock *tp = tcp_sk(sk);
196 const struct bictcp *ca = inet_csk_ca(sk);
Stephen Hemminger83803032005-06-23 12:23:25 -0700197 return max(tp->snd_cwnd, ca->last_max_cwnd);
198}
199
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300200static void bictcp_state(struct sock *sk, u8 new_state)
Stephen Hemminger83803032005-06-23 12:23:25 -0700201{
202 if (new_state == TCP_CA_Loss)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300203 bictcp_reset(inet_csk_ca(sk));
Stephen Hemminger83803032005-06-23 12:23:25 -0700204}
205
Stephen Hemminger05d05452005-12-13 23:13:13 -0800206/* Track delayed acknowledgment ratio using sliding window
Stephen Hemminger83803032005-06-23 12:23:25 -0700207 * ratio = (15*ratio + sample) / 16
208 */
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -0700209static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt)
Stephen Hemminger83803032005-06-23 12:23:25 -0700210{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300211 const struct inet_connection_sock *icsk = inet_csk(sk);
212
Stephen Hemminger05d05452005-12-13 23:13:13 -0800213 if (cnt > 0 && icsk->icsk_ca_state == TCP_CA_Open) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300214 struct bictcp *ca = inet_csk_ca(sk);
Stephen Hemminger83803032005-06-23 12:23:25 -0700215 cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
216 ca->delayed_ack += cnt;
217 }
218}
219
220
221static struct tcp_congestion_ops bictcp = {
222 .init = bictcp_init,
223 .ssthresh = bictcp_recalc_ssthresh,
224 .cong_avoid = bictcp_cong_avoid,
225 .set_state = bictcp_state,
226 .undo_cwnd = bictcp_undo_cwnd,
Stephen Hemminger83803032005-06-23 12:23:25 -0700227 .pkts_acked = bictcp_acked,
228 .owner = THIS_MODULE,
229 .name = "bic",
230};
231
232static int __init bictcp_register(void)
233{
Alexey Dobriyan65e3d722006-08-25 00:38:03 -0700234 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
Stephen Hemminger83803032005-06-23 12:23:25 -0700235 return tcp_register_congestion_control(&bictcp);
236}
237
238static void __exit bictcp_unregister(void)
239{
240 tcp_unregister_congestion_control(&bictcp);
241}
242
243module_init(bictcp_register);
244module_exit(bictcp_unregister);
245
246MODULE_AUTHOR("Stephen Hemminger");
247MODULE_LICENSE("GPL");
248MODULE_DESCRIPTION("BIC TCP");