xref: /openbmc/linux/net/ipv4/tcp_yeah.c (revision f9a82c48)
1 /*
2  *
3  *   YeAH TCP
4  *
5  * For further details look at:
6  *   https://web.archive.org/web/20080316215752/http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf
7  *
8  */
9 #include <linux/mm.h>
10 #include <linux/module.h>
11 #include <linux/skbuff.h>
12 #include <linux/inet_diag.h>
13 
14 #include <net/tcp.h>
15 
16 #include "tcp_vegas.h"
17 
18 #define TCP_YEAH_ALPHA       80 /* number of packets queued at the bottleneck */
19 #define TCP_YEAH_GAMMA        1 /* fraction of queue to be removed per rtt */
20 #define TCP_YEAH_DELTA        3 /* log minimum fraction of cwnd to be removed on loss */
21 #define TCP_YEAH_EPSILON      1 /* log maximum fraction to be removed on early decongestion */
22 #define TCP_YEAH_PHY          8 /* maximum delta from base */
23 #define TCP_YEAH_RHO         16 /* minimum number of consecutive rtt to consider competition on loss */
24 #define TCP_YEAH_ZETA        50 /* minimum number of state switches to reset reno_count */
25 
26 #define TCP_SCALABLE_AI_CNT	 100U
27 
28 /* YeAH variables */
29 struct yeah {
30 	struct vegas vegas;	/* must be first */
31 
32 	/* YeAH */
33 	u32 lastQ;
34 	u32 doing_reno_now;
35 
36 	u32 reno_count;
37 	u32 fast_count;
38 
39 	u32 pkts_acked;
40 };
41 
42 static void tcp_yeah_init(struct sock *sk)
43 {
44 	struct tcp_sock *tp = tcp_sk(sk);
45 	struct yeah *yeah = inet_csk_ca(sk);
46 
47 	tcp_vegas_init(sk);
48 
49 	yeah->doing_reno_now = 0;
50 	yeah->lastQ = 0;
51 
52 	yeah->reno_count = 2;
53 
54 	/* Ensure the MD arithmetic works.  This is somewhat pedantic,
55 	 * since I don't think we will see a cwnd this large. :) */
56 	tp->snd_cwnd_clamp = min_t(u32, tp->snd_cwnd_clamp, 0xffffffff/128);
57 }
58 
59 static void tcp_yeah_pkts_acked(struct sock *sk,
60 				const struct ack_sample *sample)
61 {
62 	const struct inet_connection_sock *icsk = inet_csk(sk);
63 	struct yeah *yeah = inet_csk_ca(sk);
64 
65 	if (icsk->icsk_ca_state == TCP_CA_Open)
66 		yeah->pkts_acked = sample->pkts_acked;
67 
68 	tcp_vegas_pkts_acked(sk, sample);
69 }
70 
71 static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 acked)
72 {
73 	struct tcp_sock *tp = tcp_sk(sk);
74 	struct yeah *yeah = inet_csk_ca(sk);
75 
76 	if (!tcp_is_cwnd_limited(sk))
77 		return;
78 
79 	if (tcp_in_slow_start(tp))
80 		tcp_slow_start(tp, acked);
81 
82 	else if (!yeah->doing_reno_now) {
83 		/* Scalable */
84 
85 		tp->snd_cwnd_cnt += yeah->pkts_acked;
86 		if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)) {
87 			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
88 				tp->snd_cwnd++;
89 			tp->snd_cwnd_cnt = 0;
90 		}
91 
92 		yeah->pkts_acked = 1;
93 
94 	} else {
95 		/* Reno */
96 		tcp_cong_avoid_ai(tp, tp->snd_cwnd, 1);
97 	}
98 
99 	/* The key players are v_vegas.beg_snd_una and v_beg_snd_nxt.
100 	 *
101 	 * These are so named because they represent the approximate values
102 	 * of snd_una and snd_nxt at the beginning of the current RTT. More
103 	 * precisely, they represent the amount of data sent during the RTT.
104 	 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
105 	 * we will calculate that (v_beg_snd_nxt - v_vegas.beg_snd_una) outstanding
106 	 * bytes of data have been ACKed during the course of the RTT, giving
107 	 * an "actual" rate of:
108 	 *
109 	 *     (v_beg_snd_nxt - v_vegas.beg_snd_una) / (rtt duration)
110 	 *
111 	 * Unfortunately, v_vegas.beg_snd_una is not exactly equal to snd_una,
112 	 * because delayed ACKs can cover more than one segment, so they
113 	 * don't line up yeahly with the boundaries of RTTs.
114 	 *
115 	 * Another unfortunate fact of life is that delayed ACKs delay the
116 	 * advance of the left edge of our send window, so that the number
117 	 * of bytes we send in an RTT is often less than our cwnd will allow.
118 	 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
119 	 */
120 
121 	if (after(ack, yeah->vegas.beg_snd_nxt)) {
122 		/* We do the Vegas calculations only if we got enough RTT
123 		 * samples that we can be reasonably sure that we got
124 		 * at least one RTT sample that wasn't from a delayed ACK.
125 		 * If we only had 2 samples total,
126 		 * then that means we're getting only 1 ACK per RTT, which
127 		 * means they're almost certainly delayed ACKs.
128 		 * If  we have 3 samples, we should be OK.
129 		 */
130 
131 		if (yeah->vegas.cntRTT > 2) {
132 			u32 rtt, queue;
133 			u64 bw;
134 
135 			/* We have enough RTT samples, so, using the Vegas
136 			 * algorithm, we determine if we should increase or
137 			 * decrease cwnd, and by how much.
138 			 */
139 
140 			/* Pluck out the RTT we are using for the Vegas
141 			 * calculations. This is the min RTT seen during the
142 			 * last RTT. Taking the min filters out the effects
143 			 * of delayed ACKs, at the cost of noticing congestion
144 			 * a bit later.
145 			 */
146 			rtt = yeah->vegas.minRTT;
147 
148 			/* Compute excess number of packets above bandwidth
149 			 * Avoid doing full 64 bit divide.
150 			 */
151 			bw = tp->snd_cwnd;
152 			bw *= rtt - yeah->vegas.baseRTT;
153 			do_div(bw, rtt);
154 			queue = bw;
155 
156 			if (queue > TCP_YEAH_ALPHA ||
157 			    rtt - yeah->vegas.baseRTT > (yeah->vegas.baseRTT / TCP_YEAH_PHY)) {
158 				if (queue > TCP_YEAH_ALPHA &&
159 				    tp->snd_cwnd > yeah->reno_count) {
160 					u32 reduction = min(queue / TCP_YEAH_GAMMA ,
161 							    tp->snd_cwnd >> TCP_YEAH_EPSILON);
162 
163 					tp->snd_cwnd -= reduction;
164 
165 					tp->snd_cwnd = max(tp->snd_cwnd,
166 							   yeah->reno_count);
167 
168 					tp->snd_ssthresh = tp->snd_cwnd;
169 				}
170 
171 				if (yeah->reno_count <= 2)
172 					yeah->reno_count = max(tp->snd_cwnd>>1, 2U);
173 				else
174 					yeah->reno_count++;
175 
176 				yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
177 							   0xffffffU);
178 			} else {
179 				yeah->fast_count++;
180 
181 				if (yeah->fast_count > TCP_YEAH_ZETA) {
182 					yeah->reno_count = 2;
183 					yeah->fast_count = 0;
184 				}
185 
186 				yeah->doing_reno_now = 0;
187 			}
188 
189 			yeah->lastQ = queue;
190 		}
191 
192 		/* Save the extent of the current window so we can use this
193 		 * at the end of the next RTT.
194 		 */
195 		yeah->vegas.beg_snd_una  = yeah->vegas.beg_snd_nxt;
196 		yeah->vegas.beg_snd_nxt  = tp->snd_nxt;
197 		yeah->vegas.beg_snd_cwnd = tp->snd_cwnd;
198 
199 		/* Wipe the slate clean for the next RTT. */
200 		yeah->vegas.cntRTT = 0;
201 		yeah->vegas.minRTT = 0x7fffffff;
202 	}
203 }
204 
205 static u32 tcp_yeah_ssthresh(struct sock *sk)
206 {
207 	const struct tcp_sock *tp = tcp_sk(sk);
208 	struct yeah *yeah = inet_csk_ca(sk);
209 	u32 reduction;
210 
211 	if (yeah->doing_reno_now < TCP_YEAH_RHO) {
212 		reduction = yeah->lastQ;
213 
214 		reduction = min(reduction, max(tp->snd_cwnd>>1, 2U));
215 
216 		reduction = max(reduction, tp->snd_cwnd >> TCP_YEAH_DELTA);
217 	} else
218 		reduction = max(tp->snd_cwnd>>1, 2U);
219 
220 	yeah->fast_count = 0;
221 	yeah->reno_count = max(yeah->reno_count>>1, 2U);
222 
223 	return max_t(int, tp->snd_cwnd - reduction, 2);
224 }
225 
226 static struct tcp_congestion_ops tcp_yeah __read_mostly = {
227 	.init		= tcp_yeah_init,
228 	.ssthresh	= tcp_yeah_ssthresh,
229 	.undo_cwnd      = tcp_reno_undo_cwnd,
230 	.cong_avoid	= tcp_yeah_cong_avoid,
231 	.set_state	= tcp_vegas_state,
232 	.cwnd_event	= tcp_vegas_cwnd_event,
233 	.get_info	= tcp_vegas_get_info,
234 	.pkts_acked	= tcp_yeah_pkts_acked,
235 
236 	.owner		= THIS_MODULE,
237 	.name		= "yeah",
238 };
239 
240 static int __init tcp_yeah_register(void)
241 {
242 	BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
243 	tcp_register_congestion_control(&tcp_yeah);
244 	return 0;
245 }
246 
247 static void __exit tcp_yeah_unregister(void)
248 {
249 	tcp_unregister_congestion_control(&tcp_yeah);
250 }
251 
252 module_init(tcp_yeah_register);
253 module_exit(tcp_yeah_unregister);
254 
255 MODULE_AUTHOR("Angelo P. Castellani");
256 MODULE_LICENSE("GPL");
257 MODULE_DESCRIPTION("YeAH TCP");
258