2 * Binary Increase Congestion control for TCP
4 * http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
5 * This is from the implementation of BICTCP in
6 * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
7 * "Binary Increase Congestion Control for Fast, Long Distance
8 * Networks" in InfoComm 2004
10 * http://netsrv.csc.ncsu.edu/export/bitcp.pdf
12 * Unless BIC is enabled and congestion window is large
13 * this behaves the same as the original Reno.
17 #include <linux/module.h>
21 #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
22 * max_cwnd = snd_cwnd * beta
26 * go to point (max+min)/N
29 static int fast_convergence
= 1;
30 static int max_increment
= 16;
31 static int low_window
= 14;
32 static int beta
= 819; /* = 819/1024 (BICTCP_BETA_SCALE) */
33 static int initial_ssthresh
;
34 static int smooth_part
= 20;
36 module_param(fast_convergence
, int, 0644);
37 MODULE_PARM_DESC(fast_convergence
, "turn on/off fast convergence");
38 module_param(max_increment
, int, 0644);
39 MODULE_PARM_DESC(max_increment
, "Limit on increment allowed during binary search");
40 module_param(low_window
, int, 0644);
41 MODULE_PARM_DESC(low_window
, "lower bound on congestion window (for TCP friendliness)");
42 module_param(beta
, int, 0644);
43 MODULE_PARM_DESC(beta
, "beta for multiplicative increase");
44 module_param(initial_ssthresh
, int, 0644);
45 MODULE_PARM_DESC(initial_ssthresh
, "initial value of slow start threshold");
46 module_param(smooth_part
, int, 0644);
47 MODULE_PARM_DESC(smooth_part
, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
50 /* BIC TCP Parameters */
52 u32 cnt
; /* increase cwnd by 1 after ACKs */
53 u32 last_max_cwnd
; /* last maximum snd_cwnd */
54 u32 loss_cwnd
; /* congestion window at last loss */
55 u32 last_cwnd
; /* the last snd_cwnd */
56 u32 last_time
; /* time when updated last_cwnd */
57 u32 epoch_start
; /* beginning of an epoch */
58 #define ACK_RATIO_SHIFT 4
59 u32 delayed_ack
; /* estimate the ratio of Packets/ACKs << 4 */
62 static inline void bictcp_reset(struct bictcp
*ca
)
65 ca
->last_max_cwnd
= 0;
69 ca
->delayed_ack
= 2 << ACK_RATIO_SHIFT
;
72 static void bictcp_init(struct sock
*sk
)
74 struct bictcp
*ca
= inet_csk_ca(sk
);
80 tcp_sk(sk
)->snd_ssthresh
= initial_ssthresh
;
84 * Compute congestion window to use.
86 static inline void bictcp_update(struct bictcp
*ca
, u32 cwnd
)
88 if (ca
->last_cwnd
== cwnd
&&
89 (s32
)(tcp_time_stamp
- ca
->last_time
) <= HZ
/ 32)
93 ca
->last_time
= tcp_time_stamp
;
95 if (ca
->epoch_start
== 0) /* record the beginning of an epoch */
96 ca
->epoch_start
= tcp_time_stamp
;
98 /* start off normal */
99 if (cwnd
<= low_window
) {
104 /* binary increase */
105 if (cwnd
< ca
->last_max_cwnd
) {
106 __u32 dist
= (ca
->last_max_cwnd
- cwnd
)
109 if (dist
> max_increment
)
110 /* linear increase */
111 ca
->cnt
= cwnd
/ max_increment
;
113 /* binary search increase */
114 ca
->cnt
= (cwnd
* smooth_part
) / BICTCP_B
;
116 /* binary search increase */
117 ca
->cnt
= cwnd
/ dist
;
119 /* slow start AMD linear increase */
120 if (cwnd
< ca
->last_max_cwnd
+ BICTCP_B
)
122 ca
->cnt
= (cwnd
* smooth_part
) / BICTCP_B
;
123 else if (cwnd
< ca
->last_max_cwnd
+ max_increment
*(BICTCP_B
-1))
125 ca
->cnt
= (cwnd
* (BICTCP_B
-1))
126 / (cwnd
- ca
->last_max_cwnd
);
128 /* linear increase */
129 ca
->cnt
= cwnd
/ max_increment
;
132 /* if in slow start or link utilization is very low */
133 if (ca
->last_max_cwnd
== 0) {
134 if (ca
->cnt
> 20) /* increase cwnd 5% per RTT */
138 ca
->cnt
= (ca
->cnt
<< ACK_RATIO_SHIFT
) / ca
->delayed_ack
;
139 if (ca
->cnt
== 0) /* cannot be zero */
143 static void bictcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
145 struct tcp_sock
*tp
= tcp_sk(sk
);
146 struct bictcp
*ca
= inet_csk_ca(sk
);
148 if (!tcp_is_cwnd_limited(sk
, in_flight
))
151 if (tp
->snd_cwnd
<= tp
->snd_ssthresh
)
154 bictcp_update(ca
, tp
->snd_cwnd
);
155 tcp_cong_avoid_ai(tp
, ca
->cnt
);
161 * behave like Reno until low_window is reached,
162 * then increase congestion window slowly
164 static u32
bictcp_recalc_ssthresh(struct sock
*sk
)
166 const struct tcp_sock
*tp
= tcp_sk(sk
);
167 struct bictcp
*ca
= inet_csk_ca(sk
);
169 ca
->epoch_start
= 0; /* end of epoch */
171 /* Wmax and fast convergence */
172 if (tp
->snd_cwnd
< ca
->last_max_cwnd
&& fast_convergence
)
173 ca
->last_max_cwnd
= (tp
->snd_cwnd
* (BICTCP_BETA_SCALE
+ beta
))
174 / (2 * BICTCP_BETA_SCALE
);
176 ca
->last_max_cwnd
= tp
->snd_cwnd
;
178 ca
->loss_cwnd
= tp
->snd_cwnd
;
181 if (tp
->snd_cwnd
<= low_window
)
182 return max(tp
->snd_cwnd
>> 1U, 2U);
184 return max((tp
->snd_cwnd
* beta
) / BICTCP_BETA_SCALE
, 2U);
187 static u32
bictcp_undo_cwnd(struct sock
*sk
)
189 const struct tcp_sock
*tp
= tcp_sk(sk
);
190 const struct bictcp
*ca
= inet_csk_ca(sk
);
191 return max(tp
->snd_cwnd
, ca
->loss_cwnd
);
194 static void bictcp_state(struct sock
*sk
, u8 new_state
)
196 if (new_state
== TCP_CA_Loss
)
197 bictcp_reset(inet_csk_ca(sk
));
200 /* Track delayed acknowledgment ratio using sliding window
201 * ratio = (15*ratio + sample) / 16
203 static void bictcp_acked(struct sock
*sk
, u32 cnt
, s32 rtt
)
205 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
207 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
208 struct bictcp
*ca
= inet_csk_ca(sk
);
209 cnt
-= ca
->delayed_ack
>> ACK_RATIO_SHIFT
;
210 ca
->delayed_ack
+= cnt
;
215 static struct tcp_congestion_ops bictcp __read_mostly
= {
217 .ssthresh
= bictcp_recalc_ssthresh
,
218 .cong_avoid
= bictcp_cong_avoid
,
219 .set_state
= bictcp_state
,
220 .undo_cwnd
= bictcp_undo_cwnd
,
221 .pkts_acked
= bictcp_acked
,
222 .owner
= THIS_MODULE
,
226 static int __init
bictcp_register(void)
228 BUILD_BUG_ON(sizeof(struct bictcp
) > ICSK_CA_PRIV_SIZE
);
229 return tcp_register_congestion_control(&bictcp
);
232 static void __exit
bictcp_unregister(void)
234 tcp_unregister_congestion_control(&bictcp
);
237 module_init(bictcp_register
);
238 module_exit(bictcp_unregister
);
240 MODULE_AUTHOR("Stephen Hemminger");
241 MODULE_LICENSE("GPL");
242 MODULE_DESCRIPTION("BIC TCP");