xref: /openbmc/linux/net/ipv4/tcp_yeah.c (revision 9a29f5fc)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *
4  *   YeAH TCP
5  *
6  * For further details look at:
7  *   https://web.archive.org/web/20080316215752/http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf
8  *
9  */
10 #include <linux/mm.h>
11 #include <linux/module.h>
12 #include <linux/skbuff.h>
13 #include <linux/inet_diag.h>
14 
15 #include <net/tcp.h>
16 
17 #include "tcp_vegas.h"
18 
19 #define TCP_YEAH_ALPHA       80 /* number of packets queued at the bottleneck */
20 #define TCP_YEAH_GAMMA        1 /* fraction of queue to be removed per rtt */
21 #define TCP_YEAH_DELTA        3 /* log minimum fraction of cwnd to be removed on loss */
22 #define TCP_YEAH_EPSILON      1 /* log maximum fraction to be removed on early decongestion */
23 #define TCP_YEAH_PHY          8 /* maximum delta from base */
24 #define TCP_YEAH_RHO         16 /* minimum number of consecutive rtt to consider competition on loss */
25 #define TCP_YEAH_ZETA        50 /* minimum number of state switches to reset reno_count */
26 
27 #define TCP_SCALABLE_AI_CNT	 100U
28 
29 /* YeAH variables */
30 struct yeah {
31 	struct vegas vegas;	/* must be first */
32 
33 	/* YeAH */
34 	u32 lastQ;
35 	u32 doing_reno_now;
36 
37 	u32 reno_count;
38 	u32 fast_count;
39 };
40 
41 static void tcp_yeah_init(struct sock *sk)
42 {
43 	struct tcp_sock *tp = tcp_sk(sk);
44 	struct yeah *yeah = inet_csk_ca(sk);
45 
46 	tcp_vegas_init(sk);
47 
48 	yeah->doing_reno_now = 0;
49 	yeah->lastQ = 0;
50 
51 	yeah->reno_count = 2;
52 
53 	/* Ensure the MD arithmetic works.  This is somewhat pedantic,
54 	 * since I don't think we will see a cwnd this large. :) */
55 	tp->snd_cwnd_clamp = min_t(u32, tp->snd_cwnd_clamp, 0xffffffff/128);
56 }
57 
58 static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 acked)
59 {
60 	struct tcp_sock *tp = tcp_sk(sk);
61 	struct yeah *yeah = inet_csk_ca(sk);
62 
63 	if (!tcp_is_cwnd_limited(sk))
64 		return;
65 
66 	if (tcp_in_slow_start(tp)) {
67 		acked = tcp_slow_start(tp, acked);
68 		if (!acked)
69 			goto do_vegas;
70 	}
71 
72 	if (!yeah->doing_reno_now) {
73 		/* Scalable */
74 		tcp_cong_avoid_ai(tp, min(tcp_snd_cwnd(tp), TCP_SCALABLE_AI_CNT),
75 				  acked);
76 	} else {
77 		/* Reno */
78 		tcp_cong_avoid_ai(tp, tcp_snd_cwnd(tp), acked);
79 	}
80 
81 	/* The key players are v_vegas.beg_snd_una and v_beg_snd_nxt.
82 	 *
83 	 * These are so named because they represent the approximate values
84 	 * of snd_una and snd_nxt at the beginning of the current RTT. More
85 	 * precisely, they represent the amount of data sent during the RTT.
86 	 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
87 	 * we will calculate that (v_beg_snd_nxt - v_vegas.beg_snd_una) outstanding
88 	 * bytes of data have been ACKed during the course of the RTT, giving
89 	 * an "actual" rate of:
90 	 *
91 	 *     (v_beg_snd_nxt - v_vegas.beg_snd_una) / (rtt duration)
92 	 *
93 	 * Unfortunately, v_vegas.beg_snd_una is not exactly equal to snd_una,
94 	 * because delayed ACKs can cover more than one segment, so they
95 	 * don't line up yeahly with the boundaries of RTTs.
96 	 *
97 	 * Another unfortunate fact of life is that delayed ACKs delay the
98 	 * advance of the left edge of our send window, so that the number
99 	 * of bytes we send in an RTT is often less than our cwnd will allow.
100 	 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
101 	 */
102 do_vegas:
103 	if (after(ack, yeah->vegas.beg_snd_nxt)) {
104 		/* We do the Vegas calculations only if we got enough RTT
105 		 * samples that we can be reasonably sure that we got
106 		 * at least one RTT sample that wasn't from a delayed ACK.
107 		 * If we only had 2 samples total,
108 		 * then that means we're getting only 1 ACK per RTT, which
109 		 * means they're almost certainly delayed ACKs.
110 		 * If  we have 3 samples, we should be OK.
111 		 */
112 
113 		if (yeah->vegas.cntRTT > 2) {
114 			u32 rtt, queue;
115 			u64 bw;
116 
117 			/* We have enough RTT samples, so, using the Vegas
118 			 * algorithm, we determine if we should increase or
119 			 * decrease cwnd, and by how much.
120 			 */
121 
122 			/* Pluck out the RTT we are using for the Vegas
123 			 * calculations. This is the min RTT seen during the
124 			 * last RTT. Taking the min filters out the effects
125 			 * of delayed ACKs, at the cost of noticing congestion
126 			 * a bit later.
127 			 */
128 			rtt = yeah->vegas.minRTT;
129 
130 			/* Compute excess number of packets above bandwidth
131 			 * Avoid doing full 64 bit divide.
132 			 */
133 			bw = tcp_snd_cwnd(tp);
134 			bw *= rtt - yeah->vegas.baseRTT;
135 			do_div(bw, rtt);
136 			queue = bw;
137 
138 			if (queue > TCP_YEAH_ALPHA ||
139 			    rtt - yeah->vegas.baseRTT > (yeah->vegas.baseRTT / TCP_YEAH_PHY)) {
140 				if (queue > TCP_YEAH_ALPHA &&
141 				    tcp_snd_cwnd(tp) > yeah->reno_count) {
142 					u32 reduction = min(queue / TCP_YEAH_GAMMA ,
143 							    tcp_snd_cwnd(tp) >> TCP_YEAH_EPSILON);
144 
145 					tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - reduction);
146 
147 					tcp_snd_cwnd_set(tp, max(tcp_snd_cwnd(tp),
148 								 yeah->reno_count));
149 
150 					tp->snd_ssthresh = tcp_snd_cwnd(tp);
151 				}
152 
153 				if (yeah->reno_count <= 2)
154 					yeah->reno_count = max(tcp_snd_cwnd(tp)>>1, 2U);
155 				else
156 					yeah->reno_count++;
157 
158 				yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
159 							   0xffffffU);
160 			} else {
161 				yeah->fast_count++;
162 
163 				if (yeah->fast_count > TCP_YEAH_ZETA) {
164 					yeah->reno_count = 2;
165 					yeah->fast_count = 0;
166 				}
167 
168 				yeah->doing_reno_now = 0;
169 			}
170 
171 			yeah->lastQ = queue;
172 		}
173 
174 		/* Save the extent of the current window so we can use this
175 		 * at the end of the next RTT.
176 		 */
177 		yeah->vegas.beg_snd_una  = yeah->vegas.beg_snd_nxt;
178 		yeah->vegas.beg_snd_nxt  = tp->snd_nxt;
179 		yeah->vegas.beg_snd_cwnd = tcp_snd_cwnd(tp);
180 
181 		/* Wipe the slate clean for the next RTT. */
182 		yeah->vegas.cntRTT = 0;
183 		yeah->vegas.minRTT = 0x7fffffff;
184 	}
185 }
186 
187 static u32 tcp_yeah_ssthresh(struct sock *sk)
188 {
189 	const struct tcp_sock *tp = tcp_sk(sk);
190 	struct yeah *yeah = inet_csk_ca(sk);
191 	u32 reduction;
192 
193 	if (yeah->doing_reno_now < TCP_YEAH_RHO) {
194 		reduction = yeah->lastQ;
195 
196 		reduction = min(reduction, max(tcp_snd_cwnd(tp)>>1, 2U));
197 
198 		reduction = max(reduction, tcp_snd_cwnd(tp) >> TCP_YEAH_DELTA);
199 	} else
200 		reduction = max(tcp_snd_cwnd(tp)>>1, 2U);
201 
202 	yeah->fast_count = 0;
203 	yeah->reno_count = max(yeah->reno_count>>1, 2U);
204 
205 	return max_t(int, tcp_snd_cwnd(tp) - reduction, 2);
206 }
207 
208 static struct tcp_congestion_ops tcp_yeah __read_mostly = {
209 	.init		= tcp_yeah_init,
210 	.ssthresh	= tcp_yeah_ssthresh,
211 	.undo_cwnd      = tcp_reno_undo_cwnd,
212 	.cong_avoid	= tcp_yeah_cong_avoid,
213 	.set_state	= tcp_vegas_state,
214 	.cwnd_event	= tcp_vegas_cwnd_event,
215 	.get_info	= tcp_vegas_get_info,
216 	.pkts_acked	= tcp_vegas_pkts_acked,
217 
218 	.owner		= THIS_MODULE,
219 	.name		= "yeah",
220 };
221 
222 static int __init tcp_yeah_register(void)
223 {
224 	BUILD_BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
225 	tcp_register_congestion_control(&tcp_yeah);
226 	return 0;
227 }
228 
229 static void __exit tcp_yeah_unregister(void)
230 {
231 	tcp_unregister_congestion_control(&tcp_yeah);
232 }
233 
234 module_init(tcp_yeah_register);
235 module_exit(tcp_yeah_unregister);
236 
237 MODULE_AUTHOR("Angelo P. Castellani");
238 MODULE_LICENSE("GPL");
239 MODULE_DESCRIPTION("YeAH TCP");
240