projects/linux·blob master · refs

241 lines · 6.7 KiB

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 */
30struct 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
41static 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
58static 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 */
102do_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 WRITE_ONCE(tp->snd_ssthresh,
151 tcp_snd_cwnd(tp));
152 }
153
154 if (yeah->reno_count <= 2)
155 yeah->reno_count = max(tcp_snd_cwnd(tp)>>1, 2U);
156 else
157 yeah->reno_count++;
158
159 yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
160 0xffffffU);
161 } else {
162 yeah->fast_count++;
163
164 if (yeah->fast_count > TCP_YEAH_ZETA) {
165 yeah->reno_count = 2;
166 yeah->fast_count = 0;
167 }
168
169 yeah->doing_reno_now = 0;
170 }
171
172 yeah->lastQ = queue;
173 }
174
175 /* Save the extent of the current window so we can use this
176 * at the end of the next RTT.
177 */
178 yeah->vegas.beg_snd_una = yeah->vegas.beg_snd_nxt;
179 yeah->vegas.beg_snd_nxt = tp->snd_nxt;
180 yeah->vegas.beg_snd_cwnd = tcp_snd_cwnd(tp);
181
182 /* Wipe the slate clean for the next RTT. */
183 yeah->vegas.cntRTT = 0;
184 yeah->vegas.minRTT = 0x7fffffff;
185 }
186}
187
188static u32 tcp_yeah_ssthresh(struct sock *sk)
189{
190 const struct tcp_sock *tp = tcp_sk(sk);
191 struct yeah *yeah = inet_csk_ca(sk);
192 u32 reduction;
193
194 if (yeah->doing_reno_now < TCP_YEAH_RHO) {
195 reduction = yeah->lastQ;
196
197 reduction = min(reduction, max(tcp_snd_cwnd(tp)>>1, 2U));
198
199 reduction = max(reduction, tcp_snd_cwnd(tp) >> TCP_YEAH_DELTA);
200 } else
201 reduction = max(tcp_snd_cwnd(tp)>>1, 2U);
202
203 yeah->fast_count = 0;
204 yeah->reno_count = max(yeah->reno_count>>1, 2U);
205
206 return max_t(int, tcp_snd_cwnd(tp) - reduction, 2);
207}
208
209static struct tcp_congestion_ops tcp_yeah __read_mostly = {
210 .init = tcp_yeah_init,
211 .ssthresh = tcp_yeah_ssthresh,
212 .undo_cwnd = tcp_reno_undo_cwnd,
213 .cong_avoid = tcp_yeah_cong_avoid,
214 .set_state = tcp_vegas_state,
215 .cwnd_event = tcp_vegas_cwnd_event,
216 .cwnd_event_tx_start = tcp_vegas_cwnd_event_tx_start,
217 .get_info = tcp_vegas_get_info,
218 .pkts_acked = tcp_vegas_pkts_acked,
219
220 .owner = THIS_MODULE,
221 .name = "yeah",
222};
223
224static int __init tcp_yeah_register(void)
225{
226 BUILD_BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
227 tcp_register_congestion_control(&tcp_yeah);
228 return 0;
229}
230
231static void __exit tcp_yeah_unregister(void)
232{
233 tcp_unregister_congestion_control(&tcp_yeah);
234}
235
236module_init(tcp_yeah_register);
237module_exit(tcp_yeah_unregister);
238
239MODULE_AUTHOR("Angelo P. Castellani");
240MODULE_LICENSE("GPL");
241MODULE_DESCRIPTION("YeAH TCP");