xref: /openbmc/linux/net/ipv4/tcp_yeah.c (revision 8684014d)
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, u32 pkts_acked, s32 rtt_us)
60 {
61 	const struct inet_connection_sock *icsk = inet_csk(sk);
62 	struct yeah *yeah = inet_csk_ca(sk);
63 
64 	if (icsk->icsk_ca_state == TCP_CA_Open)
65 		yeah->pkts_acked = pkts_acked;
66 
67 	tcp_vegas_pkts_acked(sk, pkts_acked, rtt_us);
68 }
69 
70 static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 acked)
71 {
72 	struct tcp_sock *tp = tcp_sk(sk);
73 	struct yeah *yeah = inet_csk_ca(sk);
74 
75 	if (!tcp_is_cwnd_limited(sk))
76 		return;
77 
78 	if (tp->snd_cwnd <= tp->snd_ssthresh)
79 		tcp_slow_start(tp, acked);
80 
81 	else if (!yeah->doing_reno_now) {
82 		/* Scalable */
83 
84 		tp->snd_cwnd_cnt += yeah->pkts_acked;
85 		if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)) {
86 			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
87 				tp->snd_cwnd++;
88 			tp->snd_cwnd_cnt = 0;
89 		}
90 
91 		yeah->pkts_acked = 1;
92 
93 	} else {
94 		/* Reno */
95 		tcp_cong_avoid_ai(tp, tp->snd_cwnd);
96 	}
97 
98 	/* The key players are v_vegas.beg_snd_una and v_beg_snd_nxt.
99 	 *
100 	 * These are so named because they represent the approximate values
101 	 * of snd_una and snd_nxt at the beginning of the current RTT. More
102 	 * precisely, they represent the amount of data sent during the RTT.
103 	 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
104 	 * we will calculate that (v_beg_snd_nxt - v_vegas.beg_snd_una) outstanding
105 	 * bytes of data have been ACKed during the course of the RTT, giving
106 	 * an "actual" rate of:
107 	 *
108 	 *     (v_beg_snd_nxt - v_vegas.beg_snd_una) / (rtt duration)
109 	 *
110 	 * Unfortunately, v_vegas.beg_snd_una is not exactly equal to snd_una,
111 	 * because delayed ACKs can cover more than one segment, so they
112 	 * don't line up yeahly with the boundaries of RTTs.
113 	 *
114 	 * Another unfortunate fact of life is that delayed ACKs delay the
115 	 * advance of the left edge of our send window, so that the number
116 	 * of bytes we send in an RTT is often less than our cwnd will allow.
117 	 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
118 	 */
119 
120 	if (after(ack, yeah->vegas.beg_snd_nxt)) {
121 		/* We do the Vegas calculations only if we got enough RTT
122 		 * samples that we can be reasonably sure that we got
123 		 * at least one RTT sample that wasn't from a delayed ACK.
124 		 * If we only had 2 samples total,
125 		 * then that means we're getting only 1 ACK per RTT, which
126 		 * means they're almost certainly delayed ACKs.
127 		 * If  we have 3 samples, we should be OK.
128 		 */
129 
130 		if (yeah->vegas.cntRTT > 2) {
131 			u32 rtt, queue;
132 			u64 bw;
133 
134 			/* We have enough RTT samples, so, using the Vegas
135 			 * algorithm, we determine if we should increase or
136 			 * decrease cwnd, and by how much.
137 			 */
138 
139 			/* Pluck out the RTT we are using for the Vegas
140 			 * calculations. This is the min RTT seen during the
141 			 * last RTT. Taking the min filters out the effects
142 			 * of delayed ACKs, at the cost of noticing congestion
143 			 * a bit later.
144 			 */
145 			rtt = yeah->vegas.minRTT;
146 
147 			/* Compute excess number of packets above bandwidth
148 			 * Avoid doing full 64 bit divide.
149 			 */
150 			bw = tp->snd_cwnd;
151 			bw *= rtt - yeah->vegas.baseRTT;
152 			do_div(bw, rtt);
153 			queue = bw;
154 
155 			if (queue > TCP_YEAH_ALPHA ||
156 			    rtt - yeah->vegas.baseRTT > (yeah->vegas.baseRTT / TCP_YEAH_PHY)) {
157 				if (queue > TCP_YEAH_ALPHA &&
158 				    tp->snd_cwnd > yeah->reno_count) {
159 					u32 reduction = min(queue / TCP_YEAH_GAMMA ,
160 							    tp->snd_cwnd >> TCP_YEAH_EPSILON);
161 
162 					tp->snd_cwnd -= reduction;
163 
164 					tp->snd_cwnd = max(tp->snd_cwnd,
165 							   yeah->reno_count);
166 
167 					tp->snd_ssthresh = tp->snd_cwnd;
168 				}
169 
170 				if (yeah->reno_count <= 2)
171 					yeah->reno_count = max(tp->snd_cwnd>>1, 2U);
172 				else
173 					yeah->reno_count++;
174 
175 				yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
176 							   0xffffffU);
177 			} else {
178 				yeah->fast_count++;
179 
180 				if (yeah->fast_count > TCP_YEAH_ZETA) {
181 					yeah->reno_count = 2;
182 					yeah->fast_count = 0;
183 				}
184 
185 				yeah->doing_reno_now = 0;
186 			}
187 
188 			yeah->lastQ = queue;
189 		}
190 
191 		/* Save the extent of the current window so we can use this
192 		 * at the end of the next RTT.
193 		 */
194 		yeah->vegas.beg_snd_una  = yeah->vegas.beg_snd_nxt;
195 		yeah->vegas.beg_snd_nxt  = tp->snd_nxt;
196 		yeah->vegas.beg_snd_cwnd = tp->snd_cwnd;
197 
198 		/* Wipe the slate clean for the next RTT. */
199 		yeah->vegas.cntRTT = 0;
200 		yeah->vegas.minRTT = 0x7fffffff;
201 	}
202 }
203 
204 static u32 tcp_yeah_ssthresh(struct sock *sk)
205 {
206 	const struct tcp_sock *tp = tcp_sk(sk);
207 	struct yeah *yeah = inet_csk_ca(sk);
208 	u32 reduction;
209 
210 	if (yeah->doing_reno_now < TCP_YEAH_RHO) {
211 		reduction = yeah->lastQ;
212 
213 		reduction = min(reduction, max(tp->snd_cwnd>>1, 2U));
214 
215 		reduction = max(reduction, tp->snd_cwnd >> TCP_YEAH_DELTA);
216 	} else
217 		reduction = max(tp->snd_cwnd>>1, 2U);
218 
219 	yeah->fast_count = 0;
220 	yeah->reno_count = max(yeah->reno_count>>1, 2U);
221 
222 	return tp->snd_cwnd - reduction;
223 }
224 
225 static struct tcp_congestion_ops tcp_yeah __read_mostly = {
226 	.init		= tcp_yeah_init,
227 	.ssthresh	= tcp_yeah_ssthresh,
228 	.cong_avoid	= tcp_yeah_cong_avoid,
229 	.set_state	= tcp_vegas_state,
230 	.cwnd_event	= tcp_vegas_cwnd_event,
231 	.get_info	= tcp_vegas_get_info,
232 	.pkts_acked	= tcp_yeah_pkts_acked,
233 
234 	.owner		= THIS_MODULE,
235 	.name		= "yeah",
236 };
237 
238 static int __init tcp_yeah_register(void)
239 {
240 	BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
241 	tcp_register_congestion_control(&tcp_yeah);
242 	return 0;
243 }
244 
245 static void __exit tcp_yeah_unregister(void)
246 {
247 	tcp_unregister_congestion_control(&tcp_yeah);
248 }
249 
250 module_init(tcp_yeah_register);
251 module_exit(tcp_yeah_unregister);
252 
253 MODULE_AUTHOR("Angelo P. Castellani");
254 MODULE_LICENSE("GPL");
255 MODULE_DESCRIPTION("YeAH TCP");
256