[deliverable/linux.git] / net / ipv4 / tcp_htcp.c

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

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

static int use_rtt_scaling = 1;
module_param(use_rtt_scaling, int, 0644);
MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling");

static int use_bandwidth_switch = 1;
module_param(use_bandwidth_switch, int, 0644);
MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher");

struct htcp {
	u16	alpha;		/* Fixed point arith, << 7 */
	u8	beta;           /* Fixed point arith, << 7 */
	u8	modeswitch;     /* Delay modeswitch until we had at least one congestion event */
	u32	last_cong;	/* Time since last congestion event end */
	u32	undo_last_cong;
	u16	pkts_acked;
	u32	packetcount;
	u32	minRTT;
	u32	maxRTT;

	u32	undo_maxRTT;
	u32	undo_old_maxB;

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

static inline u32 htcp_cong_time(struct htcp *ca)
{
	return jiffies - ca->last_cong;
}

static inline u32 htcp_ccount(struct htcp *ca)
{
	return htcp_cong_time(ca)/ca->minRTT;
}

static inline void htcp_reset(struct htcp *ca)
{
	ca->undo_last_cong = ca->last_cong;
	ca->undo_maxRTT = ca->maxRTT;
	ca->undo_old_maxB = ca->old_maxB;

	ca->last_cong = jiffies;
}

static u32 htcp_cwnd_undo(struct sock *sk)
{
	const struct tcp_sock *tp = tcp_sk(sk);
	struct htcp *ca = inet_csk_ca(sk);
	ca->last_cong = ca->undo_last_cong;
	ca->maxRTT = ca->undo_maxRTT;
	ca->old_maxB = ca->undo_old_maxB;
	return max(tp->snd_cwnd, (tp->snd_ssthresh<<7)/ca->beta);
}

static inline void measure_rtt(struct sock *sk)
{
	const struct inet_connection_sock *icsk = inet_csk(sk);
	const struct tcp_sock *tp = tcp_sk(sk);
	struct htcp *ca = inet_csk_ca(sk);
	u32 srtt = tp->srtt>>3;

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

	/* max RTT */
	if (icsk->icsk_ca_state == TCP_CA_Open && tp->snd_ssthresh < 0xFFFF && htcp_ccount(ca) > 3) {
		if (ca->maxRTT < ca->minRTT)
			ca->maxRTT = ca->minRTT;
		if (ca->maxRTT < srtt && srtt <= ca->maxRTT+msecs_to_jiffies(20))
			ca->maxRTT = srtt;
	}
}

static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked)
{
	const struct inet_connection_sock *icsk = inet_csk(sk);
	const struct tcp_sock *tp = tcp_sk(sk);
	struct htcp *ca = inet_csk_ca(sk);
	u32 now = tcp_time_stamp;

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

	if (!use_bandwidth_switch)
		return;

	/* achieved throughput calculations */
	if (icsk->icsk_ca_state != TCP_CA_Open &&
	    icsk->icsk_ca_state != TCP_CA_Disorder) {
		ca->packetcount = 0;
		ca->lasttime = now;
		return;
	}

	ca->packetcount += pkts_acked;

	if (ca->packetcount >= tp->snd_cwnd - (ca->alpha>>7? : 1)
			&& now - ca->lasttime >= ca->minRTT
			&& ca->minRTT > 0) {
		__u32 cur_Bi = ca->packetcount*HZ/(now - ca->lasttime);
		if (htcp_ccount(ca) <= 3) {
			/* just after backoff */
			ca->minB = ca->maxB = ca->Bi = cur_Bi;
		} else {
			ca->Bi = (3*ca->Bi + cur_Bi)/4;
			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;

		if (!between(5*maxB, 4*old_maxB, 6*old_maxB)) {
			ca->beta = BETA_MIN;
			ca->modeswitch = 0;
			return;
		}
	}

	if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
		ca->beta = (minRTT<<7)/maxRTT;
		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;
	u32 diff = htcp_cong_time(ca);

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

	if (use_rtt_scaling && minRTT) {
		u32 scale = (HZ<<3)/(10*minRTT);
		scale = min(max(scale, 1U<<2), 10U<<3); /* clamping ratio to interval [0.5,10]<<3 */
		factor = (factor<<3)/scale;
		if (!factor)
			factor = 1;
	}

	ca->alpha = 2*factor*((1<<7)-ca->beta);
	if (!ca->alpha)
		ca->alpha = ALPHA_BASE;
}

/* After we have the rtt data to calculate beta, we'd still prefer to wait one
 * 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).
 */
static void htcp_param_update(struct sock *sk)
{
	struct htcp *ca = inet_csk_ca(sk);
	u32 minRTT = ca->minRTT;
	u32 maxRTT = ca->maxRTT;

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

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

static u32 htcp_recalc_ssthresh(struct sock *sk)
{
	const struct tcp_sock *tp = tcp_sk(sk);
	const struct htcp *ca = inet_csk_ca(sk);
	htcp_param_update(sk);
	return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
}

static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
			    u32 in_flight, int data_acked)
{
	struct tcp_sock *tp = tcp_sk(sk);
	struct htcp *ca = inet_csk_ca(sk);

	if (!tcp_is_cwnd_limited(sk, in_flight))
		return;

        if (tp->snd_cwnd <= tp->snd_ssthresh)
		tcp_slow_start(tp);
	else {

		measure_rtt(sk);

		/* In dangerous area, increase slowly.
		 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
		 */
		if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
				tp->snd_cwnd++;
			tp->snd_cwnd_cnt = 0;
			htcp_alpha_update(ca);
		} else
			tp->snd_cwnd_cnt += ca->pkts_acked;

		ca->pkts_acked = 1;
	}
}

static void htcp_init(struct sock *sk)
{
	struct htcp *ca = inet_csk_ca(sk);

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

static void htcp_state(struct sock *sk, u8 new_state)
{
	switch (new_state) {
	case TCP_CA_Open:
		{
			struct htcp *ca = inet_csk_ca(sk);
			ca->last_cong = jiffies;
		}
		break;
	case TCP_CA_CWR:
	case TCP_CA_Recovery:
	case TCP_CA_Loss:
		htcp_reset(inet_csk_ca(sk));
		break;
	}
}

static struct tcp_congestion_ops htcp = {
	.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)
{
	BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
	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");
Commit	Line	Data
a7868ea6 BE	1	/*
	2	* H-TCP congestion control. The algorithm is detailed in:
	3	* R.N.Shorten, D.J.Leith:
	4	* "H-TCP: TCP for high-speed and long-distance networks"
	5	* Proc. PFLDnet, Argonne, 2004.
	6	* http://www.hamilton.ie/net/htcp3.pdf
	7	*/
	8
	9	#include <linux/config.h>
	10	#include <linux/mm.h>
	11	#include <linux/module.h>
	12	#include <net/tcp.h>
	13
	14	#define ALPHA_BASE (1<<7) /* 1.0 with shift << 7 */
	15	#define BETA_MIN (1<<6) /* 0.5 with shift << 7 */
	16	#define BETA_MAX 102 /* 0.8 with shift << 7 */
	17
	18	static int use_rtt_scaling = 1;
	19	module_param(use_rtt_scaling, int, 0644);
	20	MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling");
	21
	22	static int use_bandwidth_switch = 1;
	23	module_param(use_bandwidth_switch, int, 0644);
	24	MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher");
	25
	26	struct htcp {
	27	u16 alpha; /* Fixed point arith, << 7 */
	28	u8 beta; /* Fixed point arith, << 7 */
	29	u8 modeswitch; /* Delay modeswitch until we had at least one congestion event */
50bf3e22 BE	30	u32 last_cong; /* Time since last congestion event end */
50bf3e22 BE	31	u32 undo_last_cong;
0bc6d90b BE	32	u16 pkts_acked;
0bc6d90b BE	33	u32 packetcount;
a7868ea6 BE	34	u32 minRTT;
a7868ea6 BE	35	u32 maxRTT;
a7868ea6 BE	36
	37	u32 undo_maxRTT;
	38	u32 undo_old_maxB;
	39
	40	/* Bandwidth estimation */
	41	u32 minB;
	42	u32 maxB;
	43	u32 old_maxB;
	44	u32 Bi;
	45	u32 lasttime;
	46	};
	47
50bf3e22 BE	48	static inline u32 htcp_cong_time(struct htcp *ca)
	49	{
	50	return jiffies - ca->last_cong;
	51	}
	52
	53	static inline u32 htcp_ccount(struct htcp *ca)
	54	{
	55	return htcp_cong_time(ca)/ca->minRTT;
	56	}
	57
a7868ea6 BE	58	static inline void htcp_reset(struct htcp *ca)
a7868ea6 BE	59	{
50bf3e22	60	ca->undo_last_cong = ca->last_cong;
a7868ea6 BE	61	ca->undo_maxRTT = ca->maxRTT;
	62	ca->undo_old_maxB = ca->old_maxB;
	63
50bf3e22	64	ca->last_cong = jiffies;
a7868ea6 BE	65	}
a7868ea6 BE	66
6687e988	67	static u32 htcp_cwnd_undo(struct sock *sk)
a7868ea6	68	{
6687e988 ACM	69	const struct tcp_sock *tp = tcp_sk(sk);
6687e988 ACM	70	struct htcp *ca = inet_csk_ca(sk);
50bf3e22	71	ca->last_cong = ca->undo_last_cong;
a7868ea6 BE	72	ca->maxRTT = ca->undo_maxRTT;
	73	ca->old_maxB = ca->undo_old_maxB;
	74	return max(tp->snd_cwnd, (tp->snd_ssthresh<<7)/ca->beta);
	75	}
	76
6687e988	77	static inline void measure_rtt(struct sock *sk)
a7868ea6	78	{
6687e988 ACM	79	const struct inet_connection_sock *icsk = inet_csk(sk);
	80	const struct tcp_sock *tp = tcp_sk(sk);
	81	struct htcp *ca = inet_csk_ca(sk);
a7868ea6 BE	82	u32 srtt = tp->srtt>>3;
	83
	84	/* keep track of minimum RTT seen so far, minRTT is zero at first */
	85	if (ca->minRTT > srtt \|\| !ca->minRTT)
	86	ca->minRTT = srtt;
	87
	88	/* max RTT */
50bf3e22	89	if (icsk->icsk_ca_state == TCP_CA_Open && tp->snd_ssthresh < 0xFFFF && htcp_ccount(ca) > 3) {
a7868ea6 BE	90	if (ca->maxRTT < ca->minRTT)
a7868ea6 BE	91	ca->maxRTT = ca->minRTT;
c33ad6e4	92	if (ca->maxRTT < srtt && srtt <= ca->maxRTT+msecs_to_jiffies(20))
a7868ea6 BE	93	ca->maxRTT = srtt;
	94	}
	95	}
	96
6687e988	97	static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked)
a7868ea6	98	{
6687e988 ACM	99	const struct inet_connection_sock *icsk = inet_csk(sk);
	100	const struct tcp_sock *tp = tcp_sk(sk);
	101	struct htcp *ca = inet_csk_ca(sk);
a7868ea6 BE	102	u32 now = tcp_time_stamp;
a7868ea6 BE	103
0bc6d90b BE	104	if (icsk->icsk_ca_state == TCP_CA_Open)
	105	ca->pkts_acked = pkts_acked;
	106
	107	if (!use_bandwidth_switch)
	108	return;
	109
a7868ea6	110	/* achieved throughput calculations */
6687e988 ACM	111	if (icsk->icsk_ca_state != TCP_CA_Open &&
6687e988 ACM	112	icsk->icsk_ca_state != TCP_CA_Disorder) {
a7868ea6 BE	113	ca->packetcount = 0;
	114	ca->lasttime = now;
	115	return;
	116	}
	117
	118	ca->packetcount += pkts_acked;
	119
	120	if (ca->packetcount >= tp->snd_cwnd - (ca->alpha>>7? : 1)
	121	&& now - ca->lasttime >= ca->minRTT
	122	&& ca->minRTT > 0) {
	123	__u32 cur_Bi = ca->packetcount*HZ/(now - ca->lasttime);
50bf3e22	124	if (htcp_ccount(ca) <= 3) {
a7868ea6 BE	125	/* just after backoff */
	126	ca->minB = ca->maxB = ca->Bi = cur_Bi;
	127	} else {
	128	ca->Bi = (3*ca->Bi + cur_Bi)/4;
	129	if (ca->Bi > ca->maxB)
	130	ca->maxB = ca->Bi;
	131	if (ca->minB > ca->maxB)
	132	ca->minB = ca->maxB;
	133	}
	134	ca->packetcount = 0;
	135	ca->lasttime = now;
	136	}
	137	}
	138
	139	static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
	140	{
	141	if (use_bandwidth_switch) {
	142	u32 maxB = ca->maxB;
	143	u32 old_maxB = ca->old_maxB;
	144	ca->old_maxB = ca->maxB;
	145
	146	if (!between(5maxB, 4old_maxB, 6*old_maxB)) {
	147	ca->beta = BETA_MIN;
	148	ca->modeswitch = 0;
	149	return;
	150	}
	151	}
	152
c33ad6e4	153	if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
a7868ea6 BE	154	ca->beta = (minRTT<<7)/maxRTT;
	155	if (ca->beta < BETA_MIN)
	156	ca->beta = BETA_MIN;
	157	else if (ca->beta > BETA_MAX)
	158	ca->beta = BETA_MAX;
	159	} else {
	160	ca->beta = BETA_MIN;
	161	ca->modeswitch = 1;
	162	}
	163	}
	164
	165	static inline void htcp_alpha_update(struct htcp *ca)
	166	{
	167	u32 minRTT = ca->minRTT;
	168	u32 factor = 1;
50bf3e22	169	u32 diff = htcp_cong_time(ca);
a7868ea6 BE	170
	171	if (diff > HZ) {
	172	diff -= HZ;
	173	factor = 1+ ( 10diff + ((diff/2)(diff/2)/HZ) )/HZ;
	174	}
	175
	176	if (use_rtt_scaling && minRTT) {
	177	u32 scale = (HZ<<3)/(10*minRTT);
	178	scale = min(max(scale, 1U<<2), 10U<<3); /* clamping ratio to interval [0.5,10]<<3 */
	179	factor = (factor<<3)/scale;
	180	if (!factor)
	181	factor = 1;
	182	}
	183
	184	ca->alpha = 2factor((1<<7)-ca->beta);
	185	if (!ca->alpha)
	186	ca->alpha = ALPHA_BASE;
	187	}
	188
	189	/* After we have the rtt data to calculate beta, we'd still prefer to wait one
	190	* rtt before we adjust our beta to ensure we are working from a consistent
	191	* data.
	192	*
	193	* This function should be called when we hit a congestion event since only at
	194	* that point do we really have a real sense of maxRTT (the queues en route
	195	* were getting just too full now).
	196	*/
6687e988	197	static void htcp_param_update(struct sock *sk)
a7868ea6	198	{
6687e988	199	struct htcp *ca = inet_csk_ca(sk);
a7868ea6 BE	200	u32 minRTT = ca->minRTT;
	201	u32 maxRTT = ca->maxRTT;
	202
	203	htcp_beta_update(ca, minRTT, maxRTT);
	204	htcp_alpha_update(ca);
	205
	206	/* add slowly fading memory for maxRTT to accommodate routing changes etc */
	207	if (minRTT > 0 && maxRTT > minRTT)
	208	ca->maxRTT = minRTT + ((maxRTT-minRTT)*95)/100;
	209	}
	210
6687e988	211	static u32 htcp_recalc_ssthresh(struct sock *sk)
a7868ea6	212	{
6687e988 ACM	213	const struct tcp_sock *tp = tcp_sk(sk);
	214	const struct htcp *ca = inet_csk_ca(sk);
	215	htcp_param_update(sk);
a7868ea6 BE	216	return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
	217	}
	218
6687e988	219	static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
a7868ea6 BE	220	u32 in_flight, int data_acked)
a7868ea6 BE	221	{
6687e988 ACM	222	struct tcp_sock *tp = tcp_sk(sk);
6687e988 ACM	223	struct htcp *ca = inet_csk_ca(sk);
a7868ea6	224
f4805ede	225	if (!tcp_is_cwnd_limited(sk, in_flight))
a7868ea6 BE	226	return;
a7868ea6 BE	227
7faffa1c SH	228	if (tp->snd_cwnd <= tp->snd_ssthresh)
	229	tcp_slow_start(tp);
	230	else {
	231
6687e988	232	measure_rtt(sk);
a7868ea6	233
7faffa1c	234	/* In dangerous area, increase slowly.
a7868ea6 BE	235	* In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
a7868ea6 BE	236	*/
0bc6d90b	237	if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
a7868ea6 BE	238	if (tp->snd_cwnd < tp->snd_cwnd_clamp)
	239	tp->snd_cwnd++;
	240	tp->snd_cwnd_cnt = 0;
50bf3e22	241	htcp_alpha_update(ca);
0bc6d90b BE	242	} else
	243	tp->snd_cwnd_cnt += ca->pkts_acked;
	244
	245	ca->pkts_acked = 1;
a7868ea6 BE	246	}
	247	}
	248
6687e988	249	static void htcp_init(struct sock *sk)
a7868ea6	250	{
6687e988	251	struct htcp *ca = inet_csk_ca(sk);
a7868ea6 BE	252
	253	memset(ca, 0, sizeof(struct htcp));
	254	ca->alpha = ALPHA_BASE;
	255	ca->beta = BETA_MIN;
0bc6d90b	256	ca->pkts_acked = 1;
50bf3e22	257	ca->last_cong = jiffies;
a7868ea6 BE	258	}
a7868ea6 BE	259
6687e988	260	static void htcp_state(struct sock *sk, u8 new_state)
a7868ea6 BE	261	{
a7868ea6 BE	262	switch (new_state) {
50bf3e22 BE	263	case TCP_CA_Open:
	264	{
	265	struct htcp *ca = inet_csk_ca(sk);
	266	ca->last_cong = jiffies;
	267	}
	268	break;
a7868ea6 BE	269	case TCP_CA_CWR:
	270	case TCP_CA_Recovery:
	271	case TCP_CA_Loss:
6687e988	272	htcp_reset(inet_csk_ca(sk));
a7868ea6 BE	273	break;
	274	}
	275	}
	276
	277	static struct tcp_congestion_ops htcp = {
	278	.init = htcp_init,
	279	.ssthresh = htcp_recalc_ssthresh,
a7868ea6 BE	280	.cong_avoid = htcp_cong_avoid,
	281	.set_state = htcp_state,
	282	.undo_cwnd = htcp_cwnd_undo,
	283	.pkts_acked = measure_achieved_throughput,
	284	.owner = THIS_MODULE,
	285	.name = "htcp",
	286	};
	287
	288	static int __init htcp_register(void)
	289	{
6687e988	290	BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
a7868ea6	291	BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
a7868ea6 BE	292	return tcp_register_congestion_control(&htcp);
	293	}
	294
	295	static void __exit htcp_unregister(void)
	296	{
	297	tcp_unregister_congestion_control(&htcp);
	298	}
	299
	300	module_init(htcp_register);
	301	module_exit(htcp_unregister);
	302
	303	MODULE_AUTHOR("Baruch Even");
	304	MODULE_LICENSE("GPL");
	305	MODULE_DESCRIPTION("H-TCP");