tcp_htcp.c 7.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12
/*
 * H-TCP congestion control. The algorithm is detailed in:
 * R.N.Shorten, D.J.Leith:
 *   "H-TCP: TCP for high-speed and long-distance networks"
 *   Proc. PFLDnet, Argonne, 2004.
 * http://www.hamilton.ie/net/htcp3.pdf
 */

#include <linux/mm.h>
#include <linux/module.h>
#include <net/tcp.h>

13 14
#define ALPHA_BASE	(1<<7)	/* 1.0 with shift << 7 */
#define BETA_MIN	(1<<6)	/* 0.5 with shift << 7 */
15 16
#define BETA_MAX	102	/* 0.8 with shift << 7 */

17
static int use_rtt_scaling __read_mostly = 1;
18 19 20
module_param(use_rtt_scaling, int, 0644);
MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling");

21
static int use_bandwidth_switch __read_mostly = 1;
22 23 24 25
module_param(use_bandwidth_switch, int, 0644);
MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher");

struct htcp {
26
	u32	alpha;		/* Fixed point arith, << 7 */
27
	u8	beta;           /* Fixed point arith, << 7 */
28 29
	u8	modeswitch;	/* Delay modeswitch
				   until we had at least one congestion event */
30 31
	u16	pkts_acked;
	u32	packetcount;
32 33
	u32	minRTT;
	u32	maxRTT;
34 35
	u32	last_cong;	/* Time since last congestion event end */
	u32	undo_last_cong;
36 37 38 39 40 41 42 43 44 45 46 47

	u32	undo_maxRTT;
	u32	undo_old_maxB;

	/* Bandwidth estimation */
	u32	minB;
	u32	maxB;
	u32	old_maxB;
	u32	Bi;
	u32	lasttime;
};

48
static inline u32 htcp_cong_time(const struct htcp *ca)
49 50 51 52
{
	return jiffies - ca->last_cong;
}

53
static inline u32 htcp_ccount(const struct htcp *ca)
54
{
55
	return htcp_cong_time(ca) / ca->minRTT;
56 57
}

58 59
static inline void htcp_reset(struct htcp *ca)
{
60
	ca->undo_last_cong = ca->last_cong;
61 62 63
	ca->undo_maxRTT = ca->maxRTT;
	ca->undo_old_maxB = ca->old_maxB;

64
	ca->last_cong = jiffies;
65 66
}

67
static u32 htcp_cwnd_undo(struct sock *sk)
68
{
69 70
	const struct tcp_sock *tp = tcp_sk(sk);
	struct htcp *ca = inet_csk_ca(sk);
71

D
Doug Leith 已提交
72 73 74 75 76 77
	if (ca->undo_last_cong) {
		ca->last_cong = ca->undo_last_cong;
		ca->maxRTT = ca->undo_maxRTT;
		ca->old_maxB = ca->undo_old_maxB;
		ca->undo_last_cong = 0;
	}
78 79

	return max(tp->snd_cwnd, (tp->snd_ssthresh << 7) / ca->beta);
80 81
}

82
static inline void measure_rtt(struct sock *sk, u32 srtt)
83
{
84 85
	const struct inet_connection_sock *icsk = inet_csk(sk);
	struct htcp *ca = inet_csk_ca(sk);
86 87 88 89 90 91

	/* keep track of minimum RTT seen so far, minRTT is zero at first */
	if (ca->minRTT > srtt || !ca->minRTT)
		ca->minRTT = srtt;

	/* max RTT */
92
	if (icsk->icsk_ca_state == TCP_CA_Open) {
93 94
		if (ca->maxRTT < ca->minRTT)
			ca->maxRTT = ca->minRTT;
95 96
		if (ca->maxRTT < srtt &&
		    srtt <= ca->maxRTT + msecs_to_jiffies(20))
97 98 99 100
			ca->maxRTT = srtt;
	}
}

101
static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked, s32 rtt)
102
{
103 104 105
	const struct inet_connection_sock *icsk = inet_csk(sk);
	const struct tcp_sock *tp = tcp_sk(sk);
	struct htcp *ca = inet_csk_ca(sk);
106 107
	u32 now = tcp_time_stamp;

108 109 110
	if (icsk->icsk_ca_state == TCP_CA_Open)
		ca->pkts_acked = pkts_acked;

111 112 113
	if (rtt > 0)
		measure_rtt(sk, usecs_to_jiffies(rtt));

114 115 116
	if (!use_bandwidth_switch)
		return;

117
	/* achieved throughput calculations */
118
	if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_Disorder))) {
119 120 121 122 123 124 125
		ca->packetcount = 0;
		ca->lasttime = now;
		return;
	}

	ca->packetcount += pkts_acked;

126 127 128
	if (ca->packetcount >= tp->snd_cwnd - (ca->alpha >> 7 ? : 1) &&
	    now - ca->lasttime >= ca->minRTT &&
	    ca->minRTT > 0) {
129 130
		__u32 cur_Bi = ca->packetcount * HZ / (now - ca->lasttime);

131
		if (htcp_ccount(ca) <= 3) {
132 133 134
			/* just after backoff */
			ca->minB = ca->maxB = ca->Bi = cur_Bi;
		} else {
135
			ca->Bi = (3 * ca->Bi + cur_Bi) / 4;
136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152
			if (ca->Bi > ca->maxB)
				ca->maxB = ca->Bi;
			if (ca->minB > ca->maxB)
				ca->minB = ca->maxB;
		}
		ca->packetcount = 0;
		ca->lasttime = now;
	}
}

static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
{
	if (use_bandwidth_switch) {
		u32 maxB = ca->maxB;
		u32 old_maxB = ca->old_maxB;
		ca->old_maxB = ca->maxB;

153
		if (!between(5 * maxB, 4 * old_maxB, 6 * old_maxB)) {
154 155 156 157 158 159
			ca->beta = BETA_MIN;
			ca->modeswitch = 0;
			return;
		}
	}

B
Baruch Even 已提交
160
	if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
161
		ca->beta = (minRTT << 7) / maxRTT;
162 163 164 165 166 167 168 169 170 171 172 173 174 175
		if (ca->beta < BETA_MIN)
			ca->beta = BETA_MIN;
		else if (ca->beta > BETA_MAX)
			ca->beta = BETA_MAX;
	} else {
		ca->beta = BETA_MIN;
		ca->modeswitch = 1;
	}
}

static inline void htcp_alpha_update(struct htcp *ca)
{
	u32 minRTT = ca->minRTT;
	u32 factor = 1;
176
	u32 diff = htcp_cong_time(ca);
177 178 179

	if (diff > HZ) {
		diff -= HZ;
180
		factor = 1 + (10 * diff + ((diff / 2) * (diff / 2) / HZ)) / HZ;
181 182 183
	}

	if (use_rtt_scaling && minRTT) {
184 185 186 187 188
		u32 scale = (HZ << 3) / (10 * minRTT);

		/* clamping ratio to interval [0.5,10]<<3 */
		scale = min(max(scale, 1U << 2), 10U << 3);
		factor = (factor << 3) / scale;
189 190 191 192
		if (!factor)
			factor = 1;
	}

193
	ca->alpha = 2 * factor * ((1 << 7) - ca->beta);
194 195 196 197
	if (!ca->alpha)
		ca->alpha = ALPHA_BASE;
}

198 199
/*
 * After we have the rtt data to calculate beta, we'd still prefer to wait one
200 201 202 203 204 205 206
 * rtt before we adjust our beta to ensure we are working from a consistent
 * data.
 *
 * This function should be called when we hit a congestion event since only at
 * that point do we really have a real sense of maxRTT (the queues en route
 * were getting just too full now).
 */
207
static void htcp_param_update(struct sock *sk)
208
{
209
	struct htcp *ca = inet_csk_ca(sk);
210 211 212 213 214 215
	u32 minRTT = ca->minRTT;
	u32 maxRTT = ca->maxRTT;

	htcp_beta_update(ca, minRTT, maxRTT);
	htcp_alpha_update(ca);

216
	/* add slowly fading memory for maxRTT to accommodate routing changes */
217
	if (minRTT > 0 && maxRTT > minRTT)
218
		ca->maxRTT = minRTT + ((maxRTT - minRTT) * 95) / 100;
219 220
}

221
static u32 htcp_recalc_ssthresh(struct sock *sk)
222
{
223 224
	const struct tcp_sock *tp = tcp_sk(sk);
	const struct htcp *ca = inet_csk_ca(sk);
225

226
	htcp_param_update(sk);
227 228 229
	return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
}

230
static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
231
{
232 233
	struct tcp_sock *tp = tcp_sk(sk);
	struct htcp *ca = inet_csk_ca(sk);
234

235
	if (!tcp_is_cwnd_limited(sk))
236 237
		return;

238
	if (tp->snd_cwnd <= tp->snd_ssthresh)
239
		tcp_slow_start(tp, acked);
240 241
	else {
		/* In dangerous area, increase slowly.
242 243
		 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
		 */
244
		if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
245 246 247
			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
				tp->snd_cwnd++;
			tp->snd_cwnd_cnt = 0;
248
			htcp_alpha_update(ca);
249 250 251 252
		} else
			tp->snd_cwnd_cnt += ca->pkts_acked;

		ca->pkts_acked = 1;
253 254 255
	}
}

256
static void htcp_init(struct sock *sk)
257
{
258
	struct htcp *ca = inet_csk_ca(sk);
259 260 261 262

	memset(ca, 0, sizeof(struct htcp));
	ca->alpha = ALPHA_BASE;
	ca->beta = BETA_MIN;
263
	ca->pkts_acked = 1;
264
	ca->last_cong = jiffies;
265 266
}

267
static void htcp_state(struct sock *sk, u8 new_state)
268 269
{
	switch (new_state) {
270 271 272
	case TCP_CA_Open:
		{
			struct htcp *ca = inet_csk_ca(sk);
D
Doug Leith 已提交
273 274 275 276
			if (ca->undo_last_cong) {
				ca->last_cong = jiffies;
				ca->undo_last_cong = 0;
			}
277 278
		}
		break;
279 280 281
	case TCP_CA_CWR:
	case TCP_CA_Recovery:
	case TCP_CA_Loss:
282
		htcp_reset(inet_csk_ca(sk));
283 284 285 286
		break;
	}
}

287
static struct tcp_congestion_ops htcp __read_mostly = {
288 289 290 291 292 293 294 295 296 297 298 299
	.init		= htcp_init,
	.ssthresh	= htcp_recalc_ssthresh,
	.cong_avoid	= htcp_cong_avoid,
	.set_state	= htcp_state,
	.undo_cwnd	= htcp_cwnd_undo,
	.pkts_acked	= measure_achieved_throughput,
	.owner		= THIS_MODULE,
	.name		= "htcp",
};

static int __init htcp_register(void)
{
300
	BUILD_BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
301 302 303 304 305 306 307 308 309 310 311 312 313 314 315
	BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
	return tcp_register_congestion_control(&htcp);
}

static void __exit htcp_unregister(void)
{
	tcp_unregister_congestion_control(&htcp);
}

module_init(htcp_register);
module_exit(htcp_unregister);

MODULE_AUTHOR("Baruch Even");
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("H-TCP");