xref: /linux/net/ipv4/tcp_veno.c (revision c4dde411bc366f568dbe33366253bbfea049e8ea)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * TCP Veno congestion control
4  *
5  * This is based on the congestion detection/avoidance scheme described in
6  *    C. P. Fu, S. C. Liew.
7  *    "TCP Veno: TCP Enhancement for Transmission over Wireless Access Networks."
8  *    IEEE Journal on Selected Areas in Communication,
9  *    Feb. 2003.
10  * 	See https://www.ie.cuhk.edu.hk/fileadmin/staff_upload/soung/Journal/J3.pdf
11  */
12 
13 #include <linux/mm.h>
14 #include <linux/module.h>
15 #include <linux/skbuff.h>
16 #include <linux/inet_diag.h>
17 
18 #include <net/tcp.h>
19 
20 /* Default values of the Veno variables, in fixed-point representation
21  * with V_PARAM_SHIFT bits to the right of the binary point.
22  */
23 #define V_PARAM_SHIFT 1
24 static const int beta = 3 << V_PARAM_SHIFT;
25 
26 /* Veno variables */
27 struct veno {
28 	u8 doing_veno_now;	/* if true, do veno for this rtt */
29 	u16 cntrtt;		/* # of rtts measured within last rtt */
30 	u32 minrtt;		/* min of rtts measured within last rtt (in usec) */
31 	u32 basertt;		/* the min of all Veno rtt measurements seen (in usec) */
32 	u32 inc;		/* decide whether to increase cwnd */
33 	u32 diff;		/* calculate the diff rate */
34 };
35 
36 /* There are several situations when we must "re-start" Veno:
37  *
38  *  o when a connection is established
39  *  o after an RTO
40  *  o after fast recovery
41  *  o when we send a packet and there is no outstanding
42  *    unacknowledged data (restarting an idle connection)
43  *
44  */
45 static inline void veno_enable(struct sock *sk)
46 {
47 	struct veno *veno = inet_csk_ca(sk);
48 
49 	/* turn on Veno */
50 	veno->doing_veno_now = 1;
51 
52 	veno->minrtt = 0x7fffffff;
53 }
54 
55 static inline void veno_disable(struct sock *sk)
56 {
57 	struct veno *veno = inet_csk_ca(sk);
58 
59 	/* turn off Veno */
60 	veno->doing_veno_now = 0;
61 }
62 
63 static void tcp_veno_init(struct sock *sk)
64 {
65 	struct veno *veno = inet_csk_ca(sk);
66 
67 	veno->basertt = 0x7fffffff;
68 	veno->inc = 1;
69 	veno_enable(sk);
70 }
71 
72 /* Do rtt sampling needed for Veno. */
73 static void tcp_veno_pkts_acked(struct sock *sk,
74 				const struct ack_sample *sample)
75 {
76 	struct veno *veno = inet_csk_ca(sk);
77 	u32 vrtt;
78 
79 	if (sample->rtt_us < 0)
80 		return;
81 
82 	/* Never allow zero rtt or baseRTT */
83 	vrtt = sample->rtt_us + 1;
84 
85 	/* Filter to find propagation delay: */
86 	if (vrtt < veno->basertt)
87 		veno->basertt = vrtt;
88 
89 	/* Find the min rtt during the last rtt to find
90 	 * the current prop. delay + queuing delay:
91 	 */
92 	veno->minrtt = min(veno->minrtt, vrtt);
93 	veno->cntrtt++;
94 }
95 
96 static void tcp_veno_state(struct sock *sk, u8 ca_state)
97 {
98 	if (ca_state == TCP_CA_Open)
99 		veno_enable(sk);
100 	else
101 		veno_disable(sk);
102 }
103 
104 /*
105  * If the connection is idle and we are restarting,
106  * then we don't want to do any Veno calculations
107  * until we get fresh rtt samples.  So when we
108  * restart, we reset our Veno state to a clean
109  * state. After we get acks for this flight of
110  * packets, _then_ we can make Veno calculations
111  * again.
112  */
113 static void tcp_veno_cwnd_event(struct sock *sk, enum tcp_ca_event event)
114 {
115 	if (event == CA_EVENT_CWND_RESTART)
116 		tcp_veno_init(sk);
117 }
118 
119 static void tcp_veno_cwnd_event_tx_start(struct sock *sk)
120 {
121 	tcp_veno_init(sk);
122 }
123 
124 static void tcp_veno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
125 {
126 	struct tcp_sock *tp = tcp_sk(sk);
127 	struct veno *veno = inet_csk_ca(sk);
128 
129 	if (!veno->doing_veno_now) {
130 		tcp_reno_cong_avoid(sk, ack, acked);
131 		return;
132 	}
133 
134 	/* limited by applications */
135 	if (!tcp_is_cwnd_limited(sk))
136 		return;
137 
138 	/* We do the Veno calculations only if we got enough rtt samples */
139 	if (veno->cntrtt <= 2) {
140 		/* We don't have enough rtt samples to do the Veno
141 		 * calculation, so we'll behave like Reno.
142 		 */
143 		tcp_reno_cong_avoid(sk, ack, acked);
144 	} else {
145 		u64 target_cwnd;
146 		u32 rtt;
147 
148 		/* We have enough rtt samples, so, using the Veno
149 		 * algorithm, we determine the state of the network.
150 		 */
151 
152 		rtt = veno->minrtt;
153 
154 		target_cwnd = (u64)tcp_snd_cwnd(tp) * veno->basertt;
155 		target_cwnd <<= V_PARAM_SHIFT;
156 		do_div(target_cwnd, rtt);
157 
158 		veno->diff = (tcp_snd_cwnd(tp) << V_PARAM_SHIFT) - target_cwnd;
159 
160 		if (tcp_in_slow_start(tp)) {
161 			/* Slow start. */
162 			acked = tcp_slow_start(tp, acked);
163 			if (!acked)
164 				goto done;
165 		}
166 
167 		/* Congestion avoidance. */
168 		if (veno->diff < beta) {
169 			/* In the "non-congestive state", increase cwnd
170 			 * every rtt.
171 			 */
172 			tcp_cong_avoid_ai(tp, tcp_snd_cwnd(tp), acked);
173 		} else {
174 			/* In the "congestive state", increase cwnd
175 			 * every other rtt.
176 			 */
177 			if (tp->snd_cwnd_cnt >= tcp_snd_cwnd(tp)) {
178 				if (veno->inc &&
179 				    tcp_snd_cwnd(tp) < tp->snd_cwnd_clamp) {
180 					tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
181 					veno->inc = 0;
182 				} else
183 					veno->inc = 1;
184 				tp->snd_cwnd_cnt = 0;
185 			} else
186 				tp->snd_cwnd_cnt += acked;
187 		}
188 done:
189 		if (tcp_snd_cwnd(tp) < 2)
190 			tcp_snd_cwnd_set(tp, 2);
191 		else if (tcp_snd_cwnd(tp) > tp->snd_cwnd_clamp)
192 			tcp_snd_cwnd_set(tp, tp->snd_cwnd_clamp);
193 	}
194 	/* Wipe the slate clean for the next rtt. */
195 	/* veno->cntrtt = 0; */
196 	veno->minrtt = 0x7fffffff;
197 }
198 
199 /* Veno MD phase */
200 static u32 tcp_veno_ssthresh(struct sock *sk)
201 {
202 	const struct tcp_sock *tp = tcp_sk(sk);
203 	struct veno *veno = inet_csk_ca(sk);
204 
205 	if (veno->diff < beta)
206 		/* in "non-congestive state", cut cwnd by 1/5 */
207 		return max(tcp_snd_cwnd(tp) * 4 / 5, 2U);
208 	else
209 		/* in "congestive state", cut cwnd by 1/2 */
210 		return max(tcp_snd_cwnd(tp) >> 1U, 2U);
211 }
212 
213 static struct tcp_congestion_ops tcp_veno __read_mostly = {
214 	.init		= tcp_veno_init,
215 	.ssthresh	= tcp_veno_ssthresh,
216 	.undo_cwnd	= tcp_reno_undo_cwnd,
217 	.cong_avoid	= tcp_veno_cong_avoid,
218 	.pkts_acked	= tcp_veno_pkts_acked,
219 	.set_state	= tcp_veno_state,
220 	.cwnd_event	= tcp_veno_cwnd_event,
221 	.cwnd_event_tx_start = tcp_veno_cwnd_event_tx_start,
222 
223 	.owner		= THIS_MODULE,
224 	.name		= "veno",
225 };
226 
227 static int __init tcp_veno_register(void)
228 {
229 	BUILD_BUG_ON(sizeof(struct veno) > ICSK_CA_PRIV_SIZE);
230 	tcp_register_congestion_control(&tcp_veno);
231 	return 0;
232 }
233 
234 static void __exit tcp_veno_unregister(void)
235 {
236 	tcp_unregister_congestion_control(&tcp_veno);
237 }
238 
239 module_init(tcp_veno_register);
240 module_exit(tcp_veno_unregister);
241 
242 MODULE_AUTHOR("Bin Zhou, Cheng Peng Fu");
243 MODULE_LICENSE("GPL");
244 MODULE_DESCRIPTION("TCP Veno");
245