xref: /openbmc/linux/net/ipv4/tcp_bic.c (revision 95e9fd10)
1 /*
2  * Binary Increase Congestion control for TCP
3  * Home page:
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
9  * Available from:
10  *  http://netsrv.csc.ncsu.edu/export/bitcp.pdf
11  *
12  * Unless BIC is enabled and congestion window is large
13  * this behaves the same as the original Reno.
14  */
15 
16 #include <linux/mm.h>
17 #include <linux/module.h>
18 #include <net/tcp.h>
19 
20 
21 #define BICTCP_BETA_SCALE    1024	/* Scale factor beta calculation
22 					 * max_cwnd = snd_cwnd * beta
23 					 */
24 #define BICTCP_B		4	 /*
25 					  * In binary search,
26 					  * go to point (max+min)/N
27 					  */
28 
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;
35 
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");
48 
49 
50 /* BIC TCP Parameters */
51 struct bictcp {
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 */
60 };
61 
62 static inline void bictcp_reset(struct bictcp *ca)
63 {
64 	ca->cnt = 0;
65 	ca->last_max_cwnd = 0;
66 	ca->last_cwnd = 0;
67 	ca->last_time = 0;
68 	ca->epoch_start = 0;
69 	ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
70 }
71 
72 static void bictcp_init(struct sock *sk)
73 {
74 	struct bictcp *ca = inet_csk_ca(sk);
75 
76 	bictcp_reset(ca);
77 	ca->loss_cwnd = 0;
78 
79 	if (initial_ssthresh)
80 		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
81 }
82 
83 /*
84  * Compute congestion window to use.
85  */
86 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
87 {
88 	if (ca->last_cwnd == cwnd &&
89 	    (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
90 		return;
91 
92 	ca->last_cwnd = cwnd;
93 	ca->last_time = tcp_time_stamp;
94 
95 	if (ca->epoch_start == 0) /* record the beginning of an epoch */
96 		ca->epoch_start = tcp_time_stamp;
97 
98 	/* start off normal */
99 	if (cwnd <= low_window) {
100 		ca->cnt = cwnd;
101 		return;
102 	}
103 
104 	/* binary increase */
105 	if (cwnd < ca->last_max_cwnd) {
106 		__u32 	dist = (ca->last_max_cwnd - cwnd)
107 			/ BICTCP_B;
108 
109 		if (dist > max_increment)
110 			/* linear increase */
111 			ca->cnt = cwnd / max_increment;
112 		else if (dist <= 1U)
113 			/* binary search increase */
114 			ca->cnt = (cwnd * smooth_part) / BICTCP_B;
115 		else
116 			/* binary search increase */
117 			ca->cnt = cwnd / dist;
118 	} else {
119 		/* slow start AMD linear increase */
120 		if (cwnd < ca->last_max_cwnd + BICTCP_B)
121 			/* slow start */
122 			ca->cnt = (cwnd * smooth_part) / BICTCP_B;
123 		else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
124 			/* slow start */
125 			ca->cnt = (cwnd * (BICTCP_B-1))
126 				/ (cwnd - ca->last_max_cwnd);
127 		else
128 			/* linear increase */
129 			ca->cnt = cwnd / max_increment;
130 	}
131 
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 */
135 			ca->cnt = 20;
136 	}
137 
138 	ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
139 	if (ca->cnt == 0)			/* cannot be zero */
140 		ca->cnt = 1;
141 }
142 
143 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
144 {
145 	struct tcp_sock *tp = tcp_sk(sk);
146 	struct bictcp *ca = inet_csk_ca(sk);
147 
148 	if (!tcp_is_cwnd_limited(sk, in_flight))
149 		return;
150 
151 	if (tp->snd_cwnd <= tp->snd_ssthresh)
152 		tcp_slow_start(tp);
153 	else {
154 		bictcp_update(ca, tp->snd_cwnd);
155 		tcp_cong_avoid_ai(tp, ca->cnt);
156 	}
157 
158 }
159 
160 /*
161  *	behave like Reno until low_window is reached,
162  *	then increase congestion window slowly
163  */
164 static u32 bictcp_recalc_ssthresh(struct sock *sk)
165 {
166 	const struct tcp_sock *tp = tcp_sk(sk);
167 	struct bictcp *ca = inet_csk_ca(sk);
168 
169 	ca->epoch_start = 0;	/* end of epoch */
170 
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);
175 	else
176 		ca->last_max_cwnd = tp->snd_cwnd;
177 
178 	ca->loss_cwnd = tp->snd_cwnd;
179 
180 
181 	if (tp->snd_cwnd <= low_window)
182 		return max(tp->snd_cwnd >> 1U, 2U);
183 	else
184 		return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
185 }
186 
187 static u32 bictcp_undo_cwnd(struct sock *sk)
188 {
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);
192 }
193 
194 static void bictcp_state(struct sock *sk, u8 new_state)
195 {
196 	if (new_state == TCP_CA_Loss)
197 		bictcp_reset(inet_csk_ca(sk));
198 }
199 
200 /* Track delayed acknowledgment ratio using sliding window
201  * ratio = (15*ratio + sample) / 16
202  */
203 static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt)
204 {
205 	const struct inet_connection_sock *icsk = inet_csk(sk);
206 
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;
211 	}
212 }
213 
214 
215 static struct tcp_congestion_ops bictcp __read_mostly = {
216 	.init		= bictcp_init,
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,
223 	.name		= "bic",
224 };
225 
226 static int __init bictcp_register(void)
227 {
228 	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
229 	return tcp_register_congestion_control(&bictcp);
230 }
231 
232 static void __exit bictcp_unregister(void)
233 {
234 	tcp_unregister_congestion_control(&bictcp);
235 }
236 
237 module_init(bictcp_register);
238 module_exit(bictcp_unregister);
239 
240 MODULE_AUTHOR("Stephen Hemminger");
241 MODULE_LICENSE("GPL");
242 MODULE_DESCRIPTION("BIC TCP");
243