xref: /openbmc/linux/net/ipv4/tcp_veno.c (revision ab59859d)
1 /*
2  * TCP Veno congestion control
3  *
4  * This is based on the congestion detection/avoidance scheme described in
5  *    C. P. Fu, S. C. Liew.
6  *    "TCP Veno: TCP Enhancement for Transmission over Wireless Access Networks."
7  *    IEEE Journal on Selected Areas in Communication,
8  *    Feb. 2003.
9  * 	See http://www.ntu.edu.sg/home5/ZHOU0022/papers/CPFu03a.pdf
10  */
11 
12 #include <linux/mm.h>
13 #include <linux/module.h>
14 #include <linux/skbuff.h>
15 #include <linux/inet_diag.h>
16 
17 #include <net/tcp.h>
18 
19 /* Default values of the Veno variables, in fixed-point representation
20  * with V_PARAM_SHIFT bits to the right of the binary point.
21  */
22 #define V_PARAM_SHIFT 1
23 static const int beta = 3 << V_PARAM_SHIFT;
24 
25 /* Veno variables */
26 struct veno {
27 	u8 doing_veno_now;	/* if true, do veno for this rtt */
28 	u16 cntrtt;		/* # of rtts measured within last rtt */
29 	u32 minrtt;		/* min of rtts measured within last rtt (in usec) */
30 	u32 basertt;		/* the min of all Veno rtt measurements seen (in usec) */
31 	u32 inc;		/* decide whether to increase cwnd */
32 	u32 diff;		/* calculate the diff rate */
33 };
34 
35 /* There are several situations when we must "re-start" Veno:
36  *
37  *  o when a connection is established
38  *  o after an RTO
39  *  o after fast recovery
40  *  o when we send a packet and there is no outstanding
41  *    unacknowledged data (restarting an idle connection)
42  *
43  */
44 static inline void veno_enable(struct sock *sk)
45 {
46 	struct veno *veno = inet_csk_ca(sk);
47 
48 	/* turn on Veno */
49 	veno->doing_veno_now = 1;
50 
51 	veno->minrtt = 0x7fffffff;
52 }
53 
54 static inline void veno_disable(struct sock *sk)
55 {
56 	struct veno *veno = inet_csk_ca(sk);
57 
58 	/* turn off Veno */
59 	veno->doing_veno_now = 0;
60 }
61 
62 static void tcp_veno_init(struct sock *sk)
63 {
64 	struct veno *veno = inet_csk_ca(sk);
65 
66 	veno->basertt = 0x7fffffff;
67 	veno->inc = 1;
68 	veno_enable(sk);
69 }
70 
71 /* Do rtt sampling needed for Veno. */
72 static void tcp_veno_pkts_acked(struct sock *sk, u32 cnt, s32 rtt_us)
73 {
74 	struct veno *veno = inet_csk_ca(sk);
75 	u32 vrtt;
76 
77 	if (rtt_us < 0)
78 		return;
79 
80 	/* Never allow zero rtt or baseRTT */
81 	vrtt = rtt_us + 1;
82 
83 	/* Filter to find propagation delay: */
84 	if (vrtt < veno->basertt)
85 		veno->basertt = vrtt;
86 
87 	/* Find the min rtt during the last rtt to find
88 	 * the current prop. delay + queuing delay:
89 	 */
90 	veno->minrtt = min(veno->minrtt, vrtt);
91 	veno->cntrtt++;
92 }
93 
94 static void tcp_veno_state(struct sock *sk, u8 ca_state)
95 {
96 	if (ca_state == TCP_CA_Open)
97 		veno_enable(sk);
98 	else
99 		veno_disable(sk);
100 }
101 
102 /*
103  * If the connection is idle and we are restarting,
104  * then we don't want to do any Veno calculations
105  * until we get fresh rtt samples.  So when we
106  * restart, we reset our Veno state to a clean
107  * state. After we get acks for this flight of
108  * packets, _then_ we can make Veno calculations
109  * again.
110  */
111 static void tcp_veno_cwnd_event(struct sock *sk, enum tcp_ca_event event)
112 {
113 	if (event == CA_EVENT_CWND_RESTART || event == CA_EVENT_TX_START)
114 		tcp_veno_init(sk);
115 }
116 
117 static void tcp_veno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
118 {
119 	struct tcp_sock *tp = tcp_sk(sk);
120 	struct veno *veno = inet_csk_ca(sk);
121 
122 	if (!veno->doing_veno_now) {
123 		tcp_reno_cong_avoid(sk, ack, in_flight);
124 		return;
125 	}
126 
127 	/* limited by applications */
128 	if (!tcp_is_cwnd_limited(sk, in_flight))
129 		return;
130 
131 	/* We do the Veno calculations only if we got enough rtt samples */
132 	if (veno->cntrtt <= 2) {
133 		/* We don't have enough rtt samples to do the Veno
134 		 * calculation, so we'll behave like Reno.
135 		 */
136 		tcp_reno_cong_avoid(sk, ack, in_flight);
137 	} else {
138 		u64 target_cwnd;
139 		u32 rtt;
140 
141 		/* We have enough rtt samples, so, using the Veno
142 		 * algorithm, we determine the state of the network.
143 		 */
144 
145 		rtt = veno->minrtt;
146 
147 		target_cwnd = (tp->snd_cwnd * veno->basertt);
148 		target_cwnd <<= V_PARAM_SHIFT;
149 		do_div(target_cwnd, rtt);
150 
151 		veno->diff = (tp->snd_cwnd << V_PARAM_SHIFT) - target_cwnd;
152 
153 		if (tp->snd_cwnd <= tp->snd_ssthresh) {
154 			/* Slow start.  */
155 			tcp_slow_start(tp);
156 		} else {
157 			/* Congestion avoidance. */
158 			if (veno->diff < beta) {
159 				/* In the "non-congestive state", increase cwnd
160 				 *  every rtt.
161 				 */
162 				if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
163 					if (tp->snd_cwnd < tp->snd_cwnd_clamp)
164 						tp->snd_cwnd++;
165 					tp->snd_cwnd_cnt = 0;
166 				} else
167 					tp->snd_cwnd_cnt++;
168 			} else {
169 				/* In the "congestive state", increase cwnd
170 				 * every other rtt.
171 				 */
172 				if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
173 					if (veno->inc
174 					    && tp->snd_cwnd <
175 					    tp->snd_cwnd_clamp) {
176 						tp->snd_cwnd++;
177 						veno->inc = 0;
178 					} else
179 						veno->inc = 1;
180 					tp->snd_cwnd_cnt = 0;
181 				} else
182 					tp->snd_cwnd_cnt++;
183 			}
184 
185 		}
186 		if (tp->snd_cwnd < 2)
187 			tp->snd_cwnd = 2;
188 		else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
189 			tp->snd_cwnd = tp->snd_cwnd_clamp;
190 	}
191 	/* Wipe the slate clean for the next rtt. */
192 	/* veno->cntrtt = 0; */
193 	veno->minrtt = 0x7fffffff;
194 }
195 
196 /* Veno MD phase */
197 static u32 tcp_veno_ssthresh(struct sock *sk)
198 {
199 	const struct tcp_sock *tp = tcp_sk(sk);
200 	struct veno *veno = inet_csk_ca(sk);
201 
202 	if (veno->diff < beta)
203 		/* in "non-congestive state", cut cwnd by 1/5 */
204 		return max(tp->snd_cwnd * 4 / 5, 2U);
205 	else
206 		/* in "congestive state", cut cwnd by 1/2 */
207 		return max(tp->snd_cwnd >> 1U, 2U);
208 }
209 
210 static struct tcp_congestion_ops tcp_veno = {
211 	.flags		= TCP_CONG_RTT_STAMP,
212 	.init		= tcp_veno_init,
213 	.ssthresh	= tcp_veno_ssthresh,
214 	.cong_avoid	= tcp_veno_cong_avoid,
215 	.pkts_acked	= tcp_veno_pkts_acked,
216 	.set_state	= tcp_veno_state,
217 	.cwnd_event	= tcp_veno_cwnd_event,
218 
219 	.owner		= THIS_MODULE,
220 	.name		= "veno",
221 };
222 
223 static int __init tcp_veno_register(void)
224 {
225 	BUILD_BUG_ON(sizeof(struct veno) > ICSK_CA_PRIV_SIZE);
226 	tcp_register_congestion_control(&tcp_veno);
227 	return 0;
228 }
229 
230 static void __exit tcp_veno_unregister(void)
231 {
232 	tcp_unregister_congestion_control(&tcp_veno);
233 }
234 
235 module_init(tcp_veno_register);
236 module_exit(tcp_veno_unregister);
237 
238 MODULE_AUTHOR("Bin Zhou, Cheng Peng Fu");
239 MODULE_LICENSE("GPL");
240 MODULE_DESCRIPTION("TCP Veno");
241