xref: /openbmc/linux/net/ipv4/tcp_yeah.c (revision 37be287c)
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 
60 
61 static void tcp_yeah_pkts_acked(struct sock *sk, u32 pkts_acked, s32 rtt_us)
62 {
63 	const struct inet_connection_sock *icsk = inet_csk(sk);
64 	struct yeah *yeah = inet_csk_ca(sk);
65 
66 	if (icsk->icsk_ca_state == TCP_CA_Open)
67 		yeah->pkts_acked = pkts_acked;
68 
69 	tcp_vegas_pkts_acked(sk, pkts_acked, rtt_us);
70 }
71 
72 static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 acked,
73 				u32 in_flight)
74 {
75 	struct tcp_sock *tp = tcp_sk(sk);
76 	struct yeah *yeah = inet_csk_ca(sk);
77 
78 	if (!tcp_is_cwnd_limited(sk, in_flight))
79 		return;
80 
81 	if (tp->snd_cwnd <= tp->snd_ssthresh)
82 		tcp_slow_start(tp, acked);
83 
84 	else if (!yeah->doing_reno_now) {
85 		/* Scalable */
86 
87 		tp->snd_cwnd_cnt += yeah->pkts_acked;
88 		if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)){
89 			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
90 				tp->snd_cwnd++;
91 			tp->snd_cwnd_cnt = 0;
92 		}
93 
94 		yeah->pkts_acked = 1;
95 
96 	} else {
97 		/* Reno */
98 		tcp_cong_avoid_ai(tp, tp->snd_cwnd);
99 	}
100 
101 	/* The key players are v_vegas.beg_snd_una and v_beg_snd_nxt.
102 	 *
103 	 * These are so named because they represent the approximate values
104 	 * of snd_una and snd_nxt at the beginning of the current RTT. More
105 	 * precisely, they represent the amount of data sent during the RTT.
106 	 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
107 	 * we will calculate that (v_beg_snd_nxt - v_vegas.beg_snd_una) outstanding
108 	 * bytes of data have been ACKed during the course of the RTT, giving
109 	 * an "actual" rate of:
110 	 *
111 	 *     (v_beg_snd_nxt - v_vegas.beg_snd_una) / (rtt duration)
112 	 *
113 	 * Unfortunately, v_vegas.beg_snd_una is not exactly equal to snd_una,
114 	 * because delayed ACKs can cover more than one segment, so they
115 	 * don't line up yeahly with the boundaries of RTTs.
116 	 *
117 	 * Another unfortunate fact of life is that delayed ACKs delay the
118 	 * advance of the left edge of our send window, so that the number
119 	 * of bytes we send in an RTT is often less than our cwnd will allow.
120 	 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
121 	 */
122 
123 	if (after(ack, yeah->vegas.beg_snd_nxt)) {
124 
125 		/* We do the Vegas calculations only if we got enough RTT
126 		 * samples that we can be reasonably sure that we got
127 		 * at least one RTT sample that wasn't from a delayed ACK.
128 		 * If we only had 2 samples total,
129 		 * then that means we're getting only 1 ACK per RTT, which
130 		 * means they're almost certainly delayed ACKs.
131 		 * If  we have 3 samples, we should be OK.
132 		 */
133 
134 		if (yeah->vegas.cntRTT > 2) {
135 			u32 rtt, queue;
136 			u64 bw;
137 
138 			/* We have enough RTT samples, so, using the Vegas
139 			 * algorithm, we determine if we should increase or
140 			 * decrease cwnd, and by how much.
141 			 */
142 
143 			/* Pluck out the RTT we are using for the Vegas
144 			 * calculations. This is the min RTT seen during the
145 			 * last RTT. Taking the min filters out the effects
146 			 * of delayed ACKs, at the cost of noticing congestion
147 			 * a bit later.
148 			 */
149 			rtt = yeah->vegas.minRTT;
150 
151 			/* Compute excess number of packets above bandwidth
152 			 * Avoid doing full 64 bit divide.
153 			 */
154 			bw = tp->snd_cwnd;
155 			bw *= rtt - yeah->vegas.baseRTT;
156 			do_div(bw, rtt);
157 			queue = bw;
158 
159 			if (queue > TCP_YEAH_ALPHA ||
160 			    rtt - yeah->vegas.baseRTT > (yeah->vegas.baseRTT / TCP_YEAH_PHY)) {
161 				if (queue > TCP_YEAH_ALPHA &&
162 				    tp->snd_cwnd > yeah->reno_count) {
163 					u32 reduction = min(queue / TCP_YEAH_GAMMA ,
164 							    tp->snd_cwnd >> TCP_YEAH_EPSILON);
165 
166 					tp->snd_cwnd -= reduction;
167 
168 					tp->snd_cwnd = max(tp->snd_cwnd,
169 							   yeah->reno_count);
170 
171 					tp->snd_ssthresh = tp->snd_cwnd;
172 				}
173 
174 				if (yeah->reno_count <= 2)
175 					yeah->reno_count = max(tp->snd_cwnd>>1, 2U);
176 				else
177 					yeah->reno_count++;
178 
179 				yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
180 							   0xffffffU);
181 			} else {
182 				yeah->fast_count++;
183 
184 				if (yeah->fast_count > TCP_YEAH_ZETA) {
185 					yeah->reno_count = 2;
186 					yeah->fast_count = 0;
187 				}
188 
189 				yeah->doing_reno_now = 0;
190 			}
191 
192 			yeah->lastQ = queue;
193 
194 		}
195 
196 		/* Save the extent of the current window so we can use this
197 		 * at the end of the next RTT.
198 		 */
199 		yeah->vegas.beg_snd_una  = yeah->vegas.beg_snd_nxt;
200 		yeah->vegas.beg_snd_nxt  = tp->snd_nxt;
201 		yeah->vegas.beg_snd_cwnd = tp->snd_cwnd;
202 
203 		/* Wipe the slate clean for the next RTT. */
204 		yeah->vegas.cntRTT = 0;
205 		yeah->vegas.minRTT = 0x7fffffff;
206 	}
207 }
208 
209 static u32 tcp_yeah_ssthresh(struct sock *sk) {
210 	const struct tcp_sock *tp = tcp_sk(sk);
211 	struct yeah *yeah = inet_csk_ca(sk);
212 	u32 reduction;
213 
214 	if (yeah->doing_reno_now < TCP_YEAH_RHO) {
215 		reduction = yeah->lastQ;
216 
217 		reduction = min(reduction, max(tp->snd_cwnd>>1, 2U));
218 
219 		reduction = max(reduction, tp->snd_cwnd >> TCP_YEAH_DELTA);
220 	} else
221 		reduction = max(tp->snd_cwnd>>1, 2U);
222 
223 	yeah->fast_count = 0;
224 	yeah->reno_count = max(yeah->reno_count>>1, 2U);
225 
226 	return tp->snd_cwnd - reduction;
227 }
228 
229 static struct tcp_congestion_ops tcp_yeah __read_mostly = {
230 	.flags		= TCP_CONG_RTT_STAMP,
231 	.init		= tcp_yeah_init,
232 	.ssthresh	= tcp_yeah_ssthresh,
233 	.cong_avoid	= tcp_yeah_cong_avoid,
234 	.min_cwnd	= tcp_reno_min_cwnd,
235 	.set_state	= tcp_vegas_state,
236 	.cwnd_event	= tcp_vegas_cwnd_event,
237 	.get_info	= tcp_vegas_get_info,
238 	.pkts_acked	= tcp_yeah_pkts_acked,
239 
240 	.owner		= THIS_MODULE,
241 	.name		= "yeah",
242 };
243 
244 static int __init tcp_yeah_register(void)
245 {
246 	BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
247 	tcp_register_congestion_control(&tcp_yeah);
248 	return 0;
249 }
250 
251 static void __exit tcp_yeah_unregister(void)
252 {
253 	tcp_unregister_congestion_control(&tcp_yeah);
254 }
255 
256 module_init(tcp_yeah_register);
257 module_exit(tcp_yeah_unregister);
258 
259 MODULE_AUTHOR("Angelo P. Castellani");
260 MODULE_LICENSE("GPL");
261 MODULE_DESCRIPTION("YeAH TCP");
262