Linux Audio

Check our new training course

Loading...
v4.17
  1/*
  2 * Binary Increase Congestion control for TCP
  3 * Home page:
  4 *      http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
  5 * This is from the implementation of BICTCP in
  6 * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
  7 *  "Binary Increase Congestion Control for Fast, Long Distance
  8 *  Networks" in InfoComm 2004
  9 * Available from:
 10 *  http://netsrv.csc.ncsu.edu/export/bitcp.pdf
 11 *
 12 * Unless BIC is enabled and congestion window is large
 13 * this behaves the same as the original Reno.
 14 */
 15
 16#include <linux/mm.h>
 17#include <linux/module.h>
 18#include <net/tcp.h>
 19
 
 20#define BICTCP_BETA_SCALE    1024	/* Scale factor beta calculation
 21					 * max_cwnd = snd_cwnd * beta
 22					 */
 23#define BICTCP_B		4	 /*
 24					  * In binary search,
 25					  * go to point (max+min)/N
 26					  */
 27
 28static int fast_convergence = 1;
 29static int max_increment = 16;
 30static int low_window = 14;
 31static int beta = 819;		/* = 819/1024 (BICTCP_BETA_SCALE) */
 32static int initial_ssthresh;
 33static int smooth_part = 20;
 34
 35module_param(fast_convergence, int, 0644);
 36MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
 37module_param(max_increment, int, 0644);
 38MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
 39module_param(low_window, int, 0644);
 40MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
 41module_param(beta, int, 0644);
 42MODULE_PARM_DESC(beta, "beta for multiplicative increase");
 43module_param(initial_ssthresh, int, 0644);
 44MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
 45module_param(smooth_part, int, 0644);
 46MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
 47
 
 48/* BIC TCP Parameters */
 49struct bictcp {
 50	u32	cnt;		/* increase cwnd by 1 after ACKs */
 51	u32	last_max_cwnd;	/* last maximum snd_cwnd */
 
 52	u32	last_cwnd;	/* the last snd_cwnd */
 53	u32	last_time;	/* time when updated last_cwnd */
 54	u32	epoch_start;	/* beginning of an epoch */
 55#define ACK_RATIO_SHIFT	4
 56	u32	delayed_ack;	/* estimate the ratio of Packets/ACKs << 4 */
 57};
 58
 59static inline void bictcp_reset(struct bictcp *ca)
 60{
 61	ca->cnt = 0;
 62	ca->last_max_cwnd = 0;
 63	ca->last_cwnd = 0;
 64	ca->last_time = 0;
 65	ca->epoch_start = 0;
 66	ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
 67}
 68
 69static void bictcp_init(struct sock *sk)
 70{
 71	struct bictcp *ca = inet_csk_ca(sk);
 72
 73	bictcp_reset(ca);
 
 74
 75	if (initial_ssthresh)
 76		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
 77}
 78
 79/*
 80 * Compute congestion window to use.
 81 */
 82static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
 83{
 84	if (ca->last_cwnd == cwnd &&
 85	    (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
 86		return;
 87
 88	ca->last_cwnd = cwnd;
 89	ca->last_time = tcp_jiffies32;
 90
 91	if (ca->epoch_start == 0) /* record the beginning of an epoch */
 92		ca->epoch_start = tcp_jiffies32;
 93
 94	/* start off normal */
 95	if (cwnd <= low_window) {
 96		ca->cnt = cwnd;
 97		return;
 98	}
 99
100	/* binary increase */
101	if (cwnd < ca->last_max_cwnd) {
102		__u32	dist = (ca->last_max_cwnd - cwnd)
103			/ BICTCP_B;
104
105		if (dist > max_increment)
106			/* linear increase */
107			ca->cnt = cwnd / max_increment;
108		else if (dist <= 1U)
109			/* binary search increase */
110			ca->cnt = (cwnd * smooth_part) / BICTCP_B;
111		else
112			/* binary search increase */
113			ca->cnt = cwnd / dist;
114	} else {
115		/* slow start AMD linear increase */
116		if (cwnd < ca->last_max_cwnd + BICTCP_B)
117			/* slow start */
118			ca->cnt = (cwnd * smooth_part) / BICTCP_B;
119		else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
120			/* slow start */
121			ca->cnt = (cwnd * (BICTCP_B-1))
122				/ (cwnd - ca->last_max_cwnd);
123		else
124			/* linear increase */
125			ca->cnt = cwnd / max_increment;
126	}
127
128	/* if in slow start or link utilization is very low */
129	if (ca->last_max_cwnd == 0) {
130		if (ca->cnt > 20) /* increase cwnd 5% per RTT */
131			ca->cnt = 20;
132	}
133
134	ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
135	if (ca->cnt == 0)			/* cannot be zero */
136		ca->cnt = 1;
137}
138
139static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 
140{
141	struct tcp_sock *tp = tcp_sk(sk);
142	struct bictcp *ca = inet_csk_ca(sk);
143
144	if (!tcp_is_cwnd_limited(sk))
145		return;
146
147	if (tcp_in_slow_start(tp))
148		tcp_slow_start(tp, acked);
149	else {
150		bictcp_update(ca, tp->snd_cwnd);
151		tcp_cong_avoid_ai(tp, ca->cnt, 1);
152	}
 
153}
154
155/*
156 *	behave like Reno until low_window is reached,
157 *	then increase congestion window slowly
158 */
159static u32 bictcp_recalc_ssthresh(struct sock *sk)
160{
161	const struct tcp_sock *tp = tcp_sk(sk);
162	struct bictcp *ca = inet_csk_ca(sk);
163
164	ca->epoch_start = 0;	/* end of epoch */
165
166	/* Wmax and fast convergence */
167	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
168		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
169			/ (2 * BICTCP_BETA_SCALE);
170	else
171		ca->last_max_cwnd = tp->snd_cwnd;
172
 
 
 
173	if (tp->snd_cwnd <= low_window)
174		return max(tp->snd_cwnd >> 1U, 2U);
175	else
176		return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
177}
178
 
 
 
 
 
 
 
179static void bictcp_state(struct sock *sk, u8 new_state)
180{
181	if (new_state == TCP_CA_Loss)
182		bictcp_reset(inet_csk_ca(sk));
183}
184
185/* Track delayed acknowledgment ratio using sliding window
186 * ratio = (15*ratio + sample) / 16
187 */
188static void bictcp_acked(struct sock *sk, const struct ack_sample *sample)
189{
190	const struct inet_connection_sock *icsk = inet_csk(sk);
191
192	if (icsk->icsk_ca_state == TCP_CA_Open) {
193		struct bictcp *ca = inet_csk_ca(sk);
194
195		ca->delayed_ack += sample->pkts_acked -
196			(ca->delayed_ack >> ACK_RATIO_SHIFT);
197	}
198}
199
 
200static struct tcp_congestion_ops bictcp __read_mostly = {
201	.init		= bictcp_init,
202	.ssthresh	= bictcp_recalc_ssthresh,
203	.cong_avoid	= bictcp_cong_avoid,
204	.set_state	= bictcp_state,
205	.undo_cwnd	= tcp_reno_undo_cwnd,
206	.pkts_acked     = bictcp_acked,
207	.owner		= THIS_MODULE,
208	.name		= "bic",
209};
210
211static int __init bictcp_register(void)
212{
213	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
214	return tcp_register_congestion_control(&bictcp);
215}
216
217static void __exit bictcp_unregister(void)
218{
219	tcp_unregister_congestion_control(&bictcp);
220}
221
222module_init(bictcp_register);
223module_exit(bictcp_unregister);
224
225MODULE_AUTHOR("Stephen Hemminger");
226MODULE_LICENSE("GPL");
227MODULE_DESCRIPTION("BIC TCP");
v3.15
  1/*
  2 * Binary Increase Congestion control for TCP
  3 * Home page:
  4 *      http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
  5 * This is from the implementation of BICTCP in
  6 * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
  7 *  "Binary Increase Congestion Control for Fast, Long Distance
  8 *  Networks" in InfoComm 2004
  9 * Available from:
 10 *  http://netsrv.csc.ncsu.edu/export/bitcp.pdf
 11 *
 12 * Unless BIC is enabled and congestion window is large
 13 * this behaves the same as the original Reno.
 14 */
 15
 16#include <linux/mm.h>
 17#include <linux/module.h>
 18#include <net/tcp.h>
 19
 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
 29static int fast_convergence = 1;
 30static int max_increment = 16;
 31static int low_window = 14;
 32static int beta = 819;		/* = 819/1024 (BICTCP_BETA_SCALE) */
 33static int initial_ssthresh;
 34static int smooth_part = 20;
 35
 36module_param(fast_convergence, int, 0644);
 37MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
 38module_param(max_increment, int, 0644);
 39MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
 40module_param(low_window, int, 0644);
 41MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
 42module_param(beta, int, 0644);
 43MODULE_PARM_DESC(beta, "beta for multiplicative increase");
 44module_param(initial_ssthresh, int, 0644);
 45MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
 46module_param(smooth_part, int, 0644);
 47MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
 48
 49
 50/* BIC TCP Parameters */
 51struct bictcp {
 52	u32	cnt;		/* increase cwnd by 1 after ACKs */
 53	u32 	last_max_cwnd;	/* last maximum snd_cwnd */
 54	u32	loss_cwnd;	/* congestion window at last loss */
 55	u32	last_cwnd;	/* the last snd_cwnd */
 56	u32	last_time;	/* time when updated last_cwnd */
 57	u32	epoch_start;	/* beginning of an epoch */
 58#define ACK_RATIO_SHIFT	4
 59	u32	delayed_ack;	/* estimate the ratio of Packets/ACKs << 4 */
 60};
 61
 62static inline void bictcp_reset(struct bictcp *ca)
 63{
 64	ca->cnt = 0;
 65	ca->last_max_cwnd = 0;
 66	ca->last_cwnd = 0;
 67	ca->last_time = 0;
 68	ca->epoch_start = 0;
 69	ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
 70}
 71
 72static void bictcp_init(struct sock *sk)
 73{
 74	struct bictcp *ca = inet_csk_ca(sk);
 75
 76	bictcp_reset(ca);
 77	ca->loss_cwnd = 0;
 78
 79	if (initial_ssthresh)
 80		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
 81}
 82
 83/*
 84 * Compute congestion window to use.
 85 */
 86static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
 87{
 88	if (ca->last_cwnd == cwnd &&
 89	    (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
 90		return;
 91
 92	ca->last_cwnd = cwnd;
 93	ca->last_time = tcp_time_stamp;
 94
 95	if (ca->epoch_start == 0) /* record the beginning of an epoch */
 96		ca->epoch_start = tcp_time_stamp;
 97
 98	/* start off normal */
 99	if (cwnd <= low_window) {
100		ca->cnt = cwnd;
101		return;
102	}
103
104	/* binary increase */
105	if (cwnd < ca->last_max_cwnd) {
106		__u32 	dist = (ca->last_max_cwnd - cwnd)
107			/ BICTCP_B;
108
109		if (dist > max_increment)
110			/* linear increase */
111			ca->cnt = cwnd / max_increment;
112		else if (dist <= 1U)
113			/* binary search increase */
114			ca->cnt = (cwnd * smooth_part) / BICTCP_B;
115		else
116			/* binary search increase */
117			ca->cnt = cwnd / dist;
118	} else {
119		/* slow start AMD linear increase */
120		if (cwnd < ca->last_max_cwnd + BICTCP_B)
121			/* slow start */
122			ca->cnt = (cwnd * smooth_part) / BICTCP_B;
123		else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
124			/* slow start */
125			ca->cnt = (cwnd * (BICTCP_B-1))
126				/ (cwnd - ca->last_max_cwnd);
127		else
128			/* linear increase */
129			ca->cnt = cwnd / max_increment;
130	}
131
132	/* if in slow start or link utilization is very low */
133	if (ca->last_max_cwnd == 0) {
134		if (ca->cnt > 20) /* increase cwnd 5% per RTT */
135			ca->cnt = 20;
136	}
137
138	ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
139	if (ca->cnt == 0)			/* cannot be zero */
140		ca->cnt = 1;
141}
142
143static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked,
144			      u32 in_flight)
145{
146	struct tcp_sock *tp = tcp_sk(sk);
147	struct bictcp *ca = inet_csk_ca(sk);
148
149	if (!tcp_is_cwnd_limited(sk, in_flight))
150		return;
151
152	if (tp->snd_cwnd <= tp->snd_ssthresh)
153		tcp_slow_start(tp, acked);
154	else {
155		bictcp_update(ca, tp->snd_cwnd);
156		tcp_cong_avoid_ai(tp, ca->cnt);
157	}
158
159}
160
161/*
162 *	behave like Reno until low_window is reached,
163 *	then increase congestion window slowly
164 */
165static u32 bictcp_recalc_ssthresh(struct sock *sk)
166{
167	const struct tcp_sock *tp = tcp_sk(sk);
168	struct bictcp *ca = inet_csk_ca(sk);
169
170	ca->epoch_start = 0;	/* end of epoch */
171
172	/* Wmax and fast convergence */
173	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
174		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
175			/ (2 * BICTCP_BETA_SCALE);
176	else
177		ca->last_max_cwnd = tp->snd_cwnd;
178
179	ca->loss_cwnd = tp->snd_cwnd;
180
181
182	if (tp->snd_cwnd <= low_window)
183		return max(tp->snd_cwnd >> 1U, 2U);
184	else
185		return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
186}
187
188static u32 bictcp_undo_cwnd(struct sock *sk)
189{
190	const struct tcp_sock *tp = tcp_sk(sk);
191	const struct bictcp *ca = inet_csk_ca(sk);
192	return max(tp->snd_cwnd, ca->loss_cwnd);
193}
194
195static void bictcp_state(struct sock *sk, u8 new_state)
196{
197	if (new_state == TCP_CA_Loss)
198		bictcp_reset(inet_csk_ca(sk));
199}
200
201/* Track delayed acknowledgment ratio using sliding window
202 * ratio = (15*ratio + sample) / 16
203 */
204static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt)
205{
206	const struct inet_connection_sock *icsk = inet_csk(sk);
207
208	if (icsk->icsk_ca_state == TCP_CA_Open) {
209		struct bictcp *ca = inet_csk_ca(sk);
210		cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
211		ca->delayed_ack += cnt;
 
212	}
213}
214
215
216static struct tcp_congestion_ops bictcp __read_mostly = {
217	.init		= bictcp_init,
218	.ssthresh	= bictcp_recalc_ssthresh,
219	.cong_avoid	= bictcp_cong_avoid,
220	.set_state	= bictcp_state,
221	.undo_cwnd	= bictcp_undo_cwnd,
222	.pkts_acked     = bictcp_acked,
223	.owner		= THIS_MODULE,
224	.name		= "bic",
225};
226
227static int __init bictcp_register(void)
228{
229	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
230	return tcp_register_congestion_control(&bictcp);
231}
232
233static void __exit bictcp_unregister(void)
234{
235	tcp_unregister_congestion_control(&bictcp);
236}
237
238module_init(bictcp_register);
239module_exit(bictcp_unregister);
240
241MODULE_AUTHOR("Stephen Hemminger");
242MODULE_LICENSE("GPL");
243MODULE_DESCRIPTION("BIC TCP");