1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Binary Increase Congestion control for TCP
4 * Home page:
5 * http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
6 * This is from the implementation of BICTCP in
7 * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
8 * "Binary Increase Congestion Control for Fast, Long Distance
9 * Networks" in InfoComm 2004
10 * Available from:
11 * http://netsrv.csc.ncsu.edu/export/bitcp.pdf
12 *
13 * Unless BIC is enabled and congestion window is large
14 * this behaves the same as the original Reno.
15 */
16
17 #include <linux/mm.h>
18 #include <linux/module.h>
19 #include <net/tcp.h>
20
21 #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
22 * max_cwnd = snd_cwnd * beta
23 */
24 #define BICTCP_B 4 /*
25 * In binary search,
26 * go to point (max+min)/N
27 */
28
29 static int fast_convergence = 1;
30 static int max_increment = 16;
31 static int low_window = 14;
32 static int beta = 819; /* = 819/1024 (BICTCP_BETA_SCALE) */
33 static int initial_ssthresh;
34 static int smooth_part = 20;
35
36 module_param(fast_convergence, int, 0644);
37 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
38 module_param(max_increment, int, 0644);
39 MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
40 module_param(low_window, int, 0644);
41 MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
42 module_param(beta, int, 0644);
43 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
44 module_param(initial_ssthresh, int, 0644);
45 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
46 module_param(smooth_part, int, 0644);
47 MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
48
49 /* BIC TCP Parameters */
50 struct bictcp {
51 u32 cnt; /* increase cwnd by 1 after ACKs */
52 u32 last_max_cwnd; /* last maximum snd_cwnd */
53 u32 last_cwnd; /* the last snd_cwnd */
54 u32 last_time; /* time when updated last_cwnd */
55 u32 epoch_start; /* beginning of an epoch */
56 #define ACK_RATIO_SHIFT 4
57 u32 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */
58 };
59
bictcp_reset(struct bictcp * ca)60 static inline void bictcp_reset(struct bictcp *ca)
61 {
62 ca->cnt = 0;
63 ca->last_max_cwnd = 0;
64 ca->last_cwnd = 0;
65 ca->last_time = 0;
66 ca->epoch_start = 0;
67 ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
68 }
69
bictcp_init(struct sock * sk)70 static void bictcp_init(struct sock *sk)
71 {
72 struct bictcp *ca = inet_csk_ca(sk);
73
74 bictcp_reset(ca);
75
76 if (initial_ssthresh)
77 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
78 }
79
80 /*
81 * Compute congestion window to use.
82 */
bictcp_update(struct bictcp * ca,u32 cwnd)83 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
84 {
85 if (ca->last_cwnd == cwnd &&
86 (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
87 return;
88
89 ca->last_cwnd = cwnd;
90 ca->last_time = tcp_jiffies32;
91
92 if (ca->epoch_start == 0) /* record the beginning of an epoch */
93 ca->epoch_start = tcp_jiffies32;
94
95 /* start off normal */
96 if (cwnd <= low_window) {
97 ca->cnt = cwnd;
98 return;
99 }
100
101 /* binary increase */
102 if (cwnd < ca->last_max_cwnd) {
103 __u32 dist = (ca->last_max_cwnd - cwnd)
104 / BICTCP_B;
105
106 if (dist > max_increment)
107 /* linear increase */
108 ca->cnt = cwnd / max_increment;
109 else if (dist <= 1U)
110 /* binary search increase */
111 ca->cnt = (cwnd * smooth_part) / BICTCP_B;
112 else
113 /* binary search increase */
114 ca->cnt = cwnd / dist;
115 } else {
116 /* slow start AMD linear increase */
117 if (cwnd < ca->last_max_cwnd + BICTCP_B)
118 /* slow start */
119 ca->cnt = (cwnd * smooth_part) / BICTCP_B;
120 else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
121 /* slow start */
122 ca->cnt = (cwnd * (BICTCP_B-1))
123 / (cwnd - ca->last_max_cwnd);
124 else
125 /* linear increase */
126 ca->cnt = cwnd / max_increment;
127 }
128
129 /* if in slow start or link utilization is very low */
130 if (ca->last_max_cwnd == 0) {
131 if (ca->cnt > 20) /* increase cwnd 5% per RTT */
132 ca->cnt = 20;
133 }
134
135 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
136 if (ca->cnt == 0) /* cannot be zero */
137 ca->cnt = 1;
138 }
139
bictcp_cong_avoid(struct sock * sk,u32 ack,u32 acked)140 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
141 {
142 struct tcp_sock *tp = tcp_sk(sk);
143 struct bictcp *ca = inet_csk_ca(sk);
144
145 if (!tcp_is_cwnd_limited(sk))
146 return;
147
148 if (tcp_in_slow_start(tp)) {
149 acked = tcp_slow_start(tp, acked);
150 if (!acked)
151 return;
152 }
153 bictcp_update(ca, tcp_snd_cwnd(tp));
154 tcp_cong_avoid_ai(tp, ca->cnt, acked);
155 }
156
157 /*
158 * behave like Reno until low_window is reached,
159 * then increase congestion window slowly
160 */
bictcp_recalc_ssthresh(struct sock * sk)161 static u32 bictcp_recalc_ssthresh(struct sock *sk)
162 {
163 const struct tcp_sock *tp = tcp_sk(sk);
164 struct bictcp *ca = inet_csk_ca(sk);
165
166 ca->epoch_start = 0; /* end of epoch */
167
168 /* Wmax and fast convergence */
169 if (tcp_snd_cwnd(tp) < ca->last_max_cwnd && fast_convergence)
170 ca->last_max_cwnd = (tcp_snd_cwnd(tp) * (BICTCP_BETA_SCALE + beta))
171 / (2 * BICTCP_BETA_SCALE);
172 else
173 ca->last_max_cwnd = tcp_snd_cwnd(tp);
174
175 if (tcp_snd_cwnd(tp) <= low_window)
176 return max(tcp_snd_cwnd(tp) >> 1U, 2U);
177 else
178 return max((tcp_snd_cwnd(tp) * beta) / BICTCP_BETA_SCALE, 2U);
179 }
180
bictcp_state(struct sock * sk,u8 new_state)181 static void bictcp_state(struct sock *sk, u8 new_state)
182 {
183 if (new_state == TCP_CA_Loss)
184 bictcp_reset(inet_csk_ca(sk));
185 }
186
187 /* Track delayed acknowledgment ratio using sliding window
188 * ratio = (15*ratio + sample) / 16
189 */
bictcp_acked(struct sock * sk,const struct ack_sample * sample)190 static void bictcp_acked(struct sock *sk, const struct ack_sample *sample)
191 {
192 const struct inet_connection_sock *icsk = inet_csk(sk);
193
194 if (icsk->icsk_ca_state == TCP_CA_Open) {
195 struct bictcp *ca = inet_csk_ca(sk);
196
197 ca->delayed_ack += sample->pkts_acked -
198 (ca->delayed_ack >> ACK_RATIO_SHIFT);
199 }
200 }
201
202 static struct tcp_congestion_ops bictcp __read_mostly = {
203 .init = bictcp_init,
204 .ssthresh = bictcp_recalc_ssthresh,
205 .cong_avoid = bictcp_cong_avoid,
206 .set_state = bictcp_state,
207 .undo_cwnd = tcp_reno_undo_cwnd,
208 .pkts_acked = bictcp_acked,
209 .owner = THIS_MODULE,
210 .name = "bic",
211 };
212
bictcp_register(void)213 static int __init bictcp_register(void)
214 {
215 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
216 return tcp_register_congestion_control(&bictcp);
217 }
218
bictcp_unregister(void)219 static void __exit bictcp_unregister(void)
220 {
221 tcp_unregister_congestion_control(&bictcp);
222 }
223
224 module_init(bictcp_register);
225 module_exit(bictcp_unregister);
226
227 MODULE_AUTHOR("Stephen Hemminger");
228 MODULE_LICENSE("GPL");
229 MODULE_DESCRIPTION("BIC TCP");
230