xref: /openbmc/linux/net/ipv4/tcp_bic.c (revision 4f6cce39)
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 #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 
28 static int fast_convergence = 1;
29 static int max_increment = 16;
30 static int low_window = 14;
31 static int beta = 819;		/* = 819/1024 (BICTCP_BETA_SCALE) */
32 static int initial_ssthresh;
33 static int smooth_part = 20;
34 
35 module_param(fast_convergence, int, 0644);
36 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
37 module_param(max_increment, int, 0644);
38 MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
39 module_param(low_window, int, 0644);
40 MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
41 module_param(beta, int, 0644);
42 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
43 module_param(initial_ssthresh, int, 0644);
44 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
45 module_param(smooth_part, int, 0644);
46 MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
47 
48 /* BIC TCP Parameters */
49 struct bictcp {
50 	u32	cnt;		/* increase cwnd by 1 after ACKs */
51 	u32	last_max_cwnd;	/* last maximum snd_cwnd */
52 	u32	loss_cwnd;	/* congestion window at last loss */
53 	u32	last_cwnd;	/* the last snd_cwnd */
54 	u32	last_time;	/* time when updated last_cwnd */
55 	u32	epoch_start;	/* beginning of an epoch */
56 #define ACK_RATIO_SHIFT	4
57 	u32	delayed_ack;	/* estimate the ratio of Packets/ACKs << 4 */
58 };
59 
60 static inline void bictcp_reset(struct bictcp *ca)
61 {
62 	ca->cnt = 0;
63 	ca->last_max_cwnd = 0;
64 	ca->last_cwnd = 0;
65 	ca->last_time = 0;
66 	ca->epoch_start = 0;
67 	ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
68 }
69 
70 static void bictcp_init(struct sock *sk)
71 {
72 	struct bictcp *ca = inet_csk_ca(sk);
73 
74 	bictcp_reset(ca);
75 	ca->loss_cwnd = 0;
76 
77 	if (initial_ssthresh)
78 		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
79 }
80 
81 /*
82  * Compute congestion window to use.
83  */
84 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
85 {
86 	if (ca->last_cwnd == cwnd &&
87 	    (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
88 		return;
89 
90 	ca->last_cwnd = cwnd;
91 	ca->last_time = tcp_time_stamp;
92 
93 	if (ca->epoch_start == 0) /* record the beginning of an epoch */
94 		ca->epoch_start = tcp_time_stamp;
95 
96 	/* start off normal */
97 	if (cwnd <= low_window) {
98 		ca->cnt = cwnd;
99 		return;
100 	}
101 
102 	/* binary increase */
103 	if (cwnd < ca->last_max_cwnd) {
104 		__u32	dist = (ca->last_max_cwnd - cwnd)
105 			/ BICTCP_B;
106 
107 		if (dist > max_increment)
108 			/* linear increase */
109 			ca->cnt = cwnd / max_increment;
110 		else if (dist <= 1U)
111 			/* binary search increase */
112 			ca->cnt = (cwnd * smooth_part) / BICTCP_B;
113 		else
114 			/* binary search increase */
115 			ca->cnt = cwnd / dist;
116 	} else {
117 		/* slow start AMD linear increase */
118 		if (cwnd < ca->last_max_cwnd + BICTCP_B)
119 			/* slow start */
120 			ca->cnt = (cwnd * smooth_part) / BICTCP_B;
121 		else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
122 			/* slow start */
123 			ca->cnt = (cwnd * (BICTCP_B-1))
124 				/ (cwnd - ca->last_max_cwnd);
125 		else
126 			/* linear increase */
127 			ca->cnt = cwnd / max_increment;
128 	}
129 
130 	/* if in slow start or link utilization is very low */
131 	if (ca->last_max_cwnd == 0) {
132 		if (ca->cnt > 20) /* increase cwnd 5% per RTT */
133 			ca->cnt = 20;
134 	}
135 
136 	ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
137 	if (ca->cnt == 0)			/* cannot be zero */
138 		ca->cnt = 1;
139 }
140 
141 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
142 {
143 	struct tcp_sock *tp = tcp_sk(sk);
144 	struct bictcp *ca = inet_csk_ca(sk);
145 
146 	if (!tcp_is_cwnd_limited(sk))
147 		return;
148 
149 	if (tcp_in_slow_start(tp))
150 		tcp_slow_start(tp, acked);
151 	else {
152 		bictcp_update(ca, tp->snd_cwnd);
153 		tcp_cong_avoid_ai(tp, ca->cnt, 1);
154 	}
155 }
156 
157 /*
158  *	behave like Reno until low_window is reached,
159  *	then increase congestion window slowly
160  */
161 static u32 bictcp_recalc_ssthresh(struct sock *sk)
162 {
163 	const struct tcp_sock *tp = tcp_sk(sk);
164 	struct bictcp *ca = inet_csk_ca(sk);
165 
166 	ca->epoch_start = 0;	/* end of epoch */
167 
168 	/* Wmax and fast convergence */
169 	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
170 		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
171 			/ (2 * BICTCP_BETA_SCALE);
172 	else
173 		ca->last_max_cwnd = tp->snd_cwnd;
174 
175 	ca->loss_cwnd = tp->snd_cwnd;
176 
177 	if (tp->snd_cwnd <= low_window)
178 		return max(tp->snd_cwnd >> 1U, 2U);
179 	else
180 		return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
181 }
182 
183 static u32 bictcp_undo_cwnd(struct sock *sk)
184 {
185 	const struct tcp_sock *tp = tcp_sk(sk);
186 	const struct bictcp *ca = inet_csk_ca(sk);
187 
188 	return max(tp->snd_cwnd, ca->loss_cwnd);
189 }
190 
191 static void bictcp_state(struct sock *sk, u8 new_state)
192 {
193 	if (new_state == TCP_CA_Loss)
194 		bictcp_reset(inet_csk_ca(sk));
195 }
196 
197 /* Track delayed acknowledgment ratio using sliding window
198  * ratio = (15*ratio + sample) / 16
199  */
200 static void bictcp_acked(struct sock *sk, const struct ack_sample *sample)
201 {
202 	const struct inet_connection_sock *icsk = inet_csk(sk);
203 
204 	if (icsk->icsk_ca_state == TCP_CA_Open) {
205 		struct bictcp *ca = inet_csk_ca(sk);
206 
207 		ca->delayed_ack += sample->pkts_acked -
208 			(ca->delayed_ack >> ACK_RATIO_SHIFT);
209 	}
210 }
211 
212 static struct tcp_congestion_ops bictcp __read_mostly = {
213 	.init		= bictcp_init,
214 	.ssthresh	= bictcp_recalc_ssthresh,
215 	.cong_avoid	= bictcp_cong_avoid,
216 	.set_state	= bictcp_state,
217 	.undo_cwnd	= bictcp_undo_cwnd,
218 	.pkts_acked     = bictcp_acked,
219 	.owner		= THIS_MODULE,
220 	.name		= "bic",
221 };
222 
223 static int __init bictcp_register(void)
224 {
225 	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
226 	return tcp_register_congestion_control(&bictcp);
227 }
228 
229 static void __exit bictcp_unregister(void)
230 {
231 	tcp_unregister_congestion_control(&bictcp);
232 }
233 
234 module_init(bictcp_register);
235 module_exit(bictcp_unregister);
236 
237 MODULE_AUTHOR("Stephen Hemminger");
238 MODULE_LICENSE("GPL");
239 MODULE_DESCRIPTION("BIC TCP");
240