[thirdparty/linux.git] / net / ipv4 / tcp_dctcp.c

/* DataCenter TCP (DCTCP) congestion control.
 *
 * http://simula.stanford.edu/~alizade/Site/DCTCP.html
 *
 * This is an implementation of DCTCP over Reno, an enhancement to the
 * TCP congestion control algorithm designed for data centers. DCTCP
 * leverages Explicit Congestion Notification (ECN) in the network to
 * provide multi-bit feedback to the end hosts. DCTCP's goal is to meet
 * the following three data center transport requirements:
 *
 *  - High burst tolerance (incast due to partition/aggregate)
 *  - Low latency (short flows, queries)
 *  - High throughput (continuous data updates, large file transfers)
 *    with commodity shallow buffered switches
 *
 * The algorithm is described in detail in the following two papers:
 *
 * 1) Mohammad Alizadeh, Albert Greenberg, David A. Maltz, Jitendra Padhye,
 *    Parveen Patel, Balaji Prabhakar, Sudipta Sengupta, and Murari Sridharan:
 *      "Data Center TCP (DCTCP)", Data Center Networks session
 *      Proc. ACM SIGCOMM, New Delhi, 2010.
 *   http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf
 *
 * 2) Mohammad Alizadeh, Adel Javanmard, and Balaji Prabhakar:
 *      "Analysis of DCTCP: Stability, Convergence, and Fairness"
 *      Proc. ACM SIGMETRICS, San Jose, 2011.
 *   http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf
 *
 * Initial prototype from Abdul Kabbani, Masato Yasuda and Mohammad Alizadeh.
 *
 * Authors:
 *
 *	Daniel Borkmann <dborkman@redhat.com>
 *	Florian Westphal <fw@strlen.de>
 *	Glenn Judd <glenn.judd@morganstanley.com>
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or (at
 * your option) any later version.
 */

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

#define DCTCP_MAX_ALPHA	1024U

struct dctcp {
	u32 old_delivered;
	u32 old_delivered_ce;
	u32 prior_rcv_nxt;
	u32 dctcp_alpha;
	u32 next_seq;
	u32 ce_state;
	u32 loss_cwnd;
};

static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */
module_param(dctcp_shift_g, uint, 0644);
MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha");

static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA;
module_param(dctcp_alpha_on_init, uint, 0644);
MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value");

static struct tcp_congestion_ops dctcp_reno;

static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca)
{
	ca->next_seq = tp->snd_nxt;

	ca->old_delivered = tp->delivered;
	ca->old_delivered_ce = tp->delivered_ce;
}

static void dctcp_init(struct sock *sk)
{
	const struct tcp_sock *tp = tcp_sk(sk);

	if ((tp->ecn_flags & TCP_ECN_OK) ||
	    (sk->sk_state == TCP_LISTEN ||
	     sk->sk_state == TCP_CLOSE)) {
		struct dctcp *ca = inet_csk_ca(sk);

		ca->prior_rcv_nxt = tp->rcv_nxt;

		ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA);

		ca->loss_cwnd = 0;
		ca->ce_state = 0;

		dctcp_reset(tp, ca);
		return;
	}

	/* No ECN support? Fall back to Reno. Also need to clear
	 * ECT from sk since it is set during 3WHS for DCTCP.
	 */
	inet_csk(sk)->icsk_ca_ops = &dctcp_reno;
	INET_ECN_dontxmit(sk);
}

static u32 dctcp_ssthresh(struct sock *sk)
{
	struct dctcp *ca = inet_csk_ca(sk);
	struct tcp_sock *tp = tcp_sk(sk);

	ca->loss_cwnd = tp->snd_cwnd;
	return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->dctcp_alpha) >> 11U), 2U);
}

static void dctcp_update_alpha(struct sock *sk, u32 flags)
{
	const struct tcp_sock *tp = tcp_sk(sk);
	struct dctcp *ca = inet_csk_ca(sk);

	/* Expired RTT */
	if (!before(tp->snd_una, ca->next_seq)) {
		u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce;
		u32 alpha = ca->dctcp_alpha;

		/* alpha = (1 - g) * alpha + g * F */

		alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
		if (delivered_ce) {
			u32 delivered = tp->delivered - ca->old_delivered;

			/* If dctcp_shift_g == 1, a 32bit value would overflow
			 * after 8 M packets.
			 */
			delivered_ce <<= (10 - dctcp_shift_g);
			delivered_ce /= max(1U, delivered);

			alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA);
		}
		/* dctcp_alpha can be read from dctcp_get_info() without
		 * synchro, so we ask compiler to not use dctcp_alpha
		 * as a temporary variable in prior operations.
		 */
		WRITE_ONCE(ca->dctcp_alpha, alpha);
		dctcp_reset(tp, ca);
	}
}

static void dctcp_react_to_loss(struct sock *sk)
{
	struct dctcp *ca = inet_csk_ca(sk);
	struct tcp_sock *tp = tcp_sk(sk);

	ca->loss_cwnd = tp->snd_cwnd;
	tp->snd_ssthresh = max(tp->snd_cwnd >> 1U, 2U);
}

static void dctcp_state(struct sock *sk, u8 new_state)
{
	if (new_state == TCP_CA_Recovery &&
	    new_state != inet_csk(sk)->icsk_ca_state)
		dctcp_react_to_loss(sk);
	/* We handle RTO in dctcp_cwnd_event to ensure that we perform only
	 * one loss-adjustment per RTT.
	 */
}

static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
{
	struct dctcp *ca = inet_csk_ca(sk);

	switch (ev) {
	case CA_EVENT_ECN_IS_CE:
	case CA_EVENT_ECN_NO_CE:
		dctcp_ece_ack_update(sk, ev, &ca->prior_rcv_nxt, &ca->ce_state);
		break;
	case CA_EVENT_LOSS:
		dctcp_react_to_loss(sk);
		break;
	default:
		/* Don't care for the rest. */
		break;
	}
}

static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
			     union tcp_cc_info *info)
{
	const struct dctcp *ca = inet_csk_ca(sk);
	const struct tcp_sock *tp = tcp_sk(sk);

	/* Fill it also in case of VEGASINFO due to req struct limits.
	 * We can still correctly retrieve it later.
	 */
	if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) ||
	    ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
		memset(&info->dctcp, 0, sizeof(info->dctcp));
		if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) {
			info->dctcp.dctcp_enabled = 1;
			info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
			info->dctcp.dctcp_alpha = ca->dctcp_alpha;
			info->dctcp.dctcp_ab_ecn = tp->mss_cache *
						   (tp->delivered_ce - ca->old_delivered_ce);
			info->dctcp.dctcp_ab_tot = tp->mss_cache *
						   (tp->delivered - ca->old_delivered);
		}

		*attr = INET_DIAG_DCTCPINFO;
		return sizeof(info->dctcp);
	}
	return 0;
}

static u32 dctcp_cwnd_undo(struct sock *sk)
{
	const struct dctcp *ca = inet_csk_ca(sk);

	return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
}

static struct tcp_congestion_ops dctcp __read_mostly = {
	.init		= dctcp_init,
	.in_ack_event   = dctcp_update_alpha,
	.cwnd_event	= dctcp_cwnd_event,
	.ssthresh	= dctcp_ssthresh,
	.cong_avoid	= tcp_reno_cong_avoid,
	.undo_cwnd	= dctcp_cwnd_undo,
	.set_state	= dctcp_state,
	.get_info	= dctcp_get_info,
	.flags		= TCP_CONG_NEEDS_ECN,
	.owner		= THIS_MODULE,
	.name		= "dctcp",
};

static struct tcp_congestion_ops dctcp_reno __read_mostly = {
	.ssthresh	= tcp_reno_ssthresh,
	.cong_avoid	= tcp_reno_cong_avoid,
	.undo_cwnd	= tcp_reno_undo_cwnd,
	.get_info	= dctcp_get_info,
	.owner		= THIS_MODULE,
	.name		= "dctcp-reno",
};

static int __init dctcp_register(void)
{
	BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE);
	return tcp_register_congestion_control(&dctcp);
}

static void __exit dctcp_unregister(void)
{
	tcp_unregister_congestion_control(&dctcp);
}

module_init(dctcp_register);
module_exit(dctcp_unregister);

MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");

MODULE_LICENSE("GPL v2");
MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");
Commit	Line	Data
e3118e83 DB	1	/* DataCenter TCP (DCTCP) congestion control.
	2	*
	3	* http://simula.stanford.edu/~alizade/Site/DCTCP.html
	4	*
	5	* This is an implementation of DCTCP over Reno, an enhancement to the
	6	* TCP congestion control algorithm designed for data centers. DCTCP
	7	* leverages Explicit Congestion Notification (ECN) in the network to
	8	* provide multi-bit feedback to the end hosts. DCTCP's goal is to meet
	9	* the following three data center transport requirements:
	10	*
	11	* - High burst tolerance (incast due to partition/aggregate)
	12	* - Low latency (short flows, queries)
	13	* - High throughput (continuous data updates, large file transfers)
	14	* with commodity shallow buffered switches
	15	*
	16	* The algorithm is described in detail in the following two papers:
	17	*
	18	* 1) Mohammad Alizadeh, Albert Greenberg, David A. Maltz, Jitendra Padhye,
	19	* Parveen Patel, Balaji Prabhakar, Sudipta Sengupta, and Murari Sridharan:
	20	* "Data Center TCP (DCTCP)", Data Center Networks session
	21	* Proc. ACM SIGCOMM, New Delhi, 2010.
	22	* http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf
	23	*
	24	* 2) Mohammad Alizadeh, Adel Javanmard, and Balaji Prabhakar:
	25	* "Analysis of DCTCP: Stability, Convergence, and Fairness"
	26	* Proc. ACM SIGMETRICS, San Jose, 2011.
	27	* http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf
	28	*
	29	* Initial prototype from Abdul Kabbani, Masato Yasuda and Mohammad Alizadeh.
	30	*
	31	* Authors:
	32	*
	33	* Daniel Borkmann <dborkman@redhat.com>
	34	* Florian Westphal <fw@strlen.de>
	35	* Glenn Judd <glenn.judd@morganstanley.com>
	36	*
	37	* This program is free software; you can redistribute it and/or modify
	38	* it under the terms of the GNU General Public License as published by
	39	* the Free Software Foundation; either version 2 of the License, or (at
	40	* your option) any later version.
	41	*/
	42
	43	#include <linux/module.h>
	44	#include <linux/mm.h>
	45	#include <net/tcp.h>
	46	#include <linux/inet_diag.h>
ffd177de	47	#include "tcp_dctcp.h"
e3118e83 DB	48
	49	#define DCTCP_MAX_ALPHA 1024U
	50
	51	struct dctcp {
e3058450 ED	52	u32 old_delivered;
e3058450 ED	53	u32 old_delivered_ce;
e3118e83 DB	54	u32 prior_rcv_nxt;
	55	u32 dctcp_alpha;
	56	u32 next_seq;
	57	u32 ce_state;
ce6dd233	58	u32 loss_cwnd;
e3118e83 DB	59	};
	60
	61	static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */
	62	module_param(dctcp_shift_g, uint, 0644);
	63	MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha");
	64
	65	static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA;
	66	module_param(dctcp_alpha_on_init, uint, 0644);
	67	MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value");
	68
e3118e83 DB	69	static struct tcp_congestion_ops dctcp_reno;
	70
	71	static void dctcp_reset(const struct tcp_sock tp, struct dctcp ca)
	72	{
	73	ca->next_seq = tp->snd_nxt;
	74
e3058450 ED	75	ca->old_delivered = tp->delivered;
e3058450 ED	76	ca->old_delivered_ce = tp->delivered_ce;
e3118e83 DB	77	}
	78
	79	static void dctcp_init(struct sock *sk)
	80	{
	81	const struct tcp_sock *tp = tcp_sk(sk);
	82
	83	if ((tp->ecn_flags & TCP_ECN_OK) \|\|
	84	(sk->sk_state == TCP_LISTEN \|\|
	85	sk->sk_state == TCP_CLOSE)) {
	86	struct dctcp *ca = inet_csk_ca(sk);
	87
e3118e83 DB	88	ca->prior_rcv_nxt = tp->rcv_nxt;
	89
	90	ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA);
	91
ce6dd233	92	ca->loss_cwnd = 0;
e3118e83 DB	93	ca->ce_state = 0;
	94
	95	dctcp_reset(tp, ca);
	96	return;
	97	}
	98
	99	/* No ECN support? Fall back to Reno. Also need to clear
	100	* ECT from sk since it is set during 3WHS for DCTCP.
	101	*/
	102	inet_csk(sk)->icsk_ca_ops = &dctcp_reno;
	103	INET_ECN_dontxmit(sk);
	104	}
	105
	106	static u32 dctcp_ssthresh(struct sock *sk)
	107	{
ce6dd233	108	struct dctcp *ca = inet_csk_ca(sk);
e3118e83 DB	109	struct tcp_sock *tp = tcp_sk(sk);
e3118e83 DB	110
ce6dd233	111	ca->loss_cwnd = tp->snd_cwnd;
e3118e83 DB	112	return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->dctcp_alpha) >> 11U), 2U);
	113	}
	114
e3118e83 DB	115	static void dctcp_update_alpha(struct sock *sk, u32 flags)
e3118e83 DB	116	{
343dfaa1	117	const struct tcp_sock *tp = tcp_sk(sk);
e3118e83	118	struct dctcp *ca = inet_csk_ca(sk);
e3118e83 DB	119
	120	/* Expired RTT */
	121	if (!before(tp->snd_una, ca->next_seq)) {
e3058450	122	u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce;
f9c2ff22	123	u32 alpha = ca->dctcp_alpha;
e3118e83 DB	124
e3118e83 DB	125	/* alpha = (1 - g) * alpha + g * F */
e3118e83	126
c80dbe04	127	alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
e3058450 ED	128	if (delivered_ce) {
	129	u32 delivered = tp->delivered - ca->old_delivered;
	130
f9c2ff22	131	/* If dctcp_shift_g == 1, a 32bit value would overflow
e3058450	132	* after 8 M packets.
f9c2ff22	133	*/
e3058450 ED	134	delivered_ce <<= (10 - dctcp_shift_g);
e3058450 ED	135	delivered_ce /= max(1U, delivered);
f9c2ff22	136
e3058450	137	alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA);
f9c2ff22 ED	138	}
	139	/* dctcp_alpha can be read from dctcp_get_info() without
	140	* synchro, so we ask compiler to not use dctcp_alpha
	141	* as a temporary variable in prior operations.
	142	*/
	143	WRITE_ONCE(ca->dctcp_alpha, alpha);
e3118e83 DB	144	dctcp_reset(tp, ca);
	145	}
	146	}
	147
aecfde23	148	static void dctcp_react_to_loss(struct sock *sk)
e3118e83	149	{
aecfde23 KDS	150	struct dctcp *ca = inet_csk_ca(sk);
aecfde23 KDS	151	struct tcp_sock *tp = tcp_sk(sk);
e3118e83	152
aecfde23 KDS	153	ca->loss_cwnd = tp->snd_cwnd;
	154	tp->snd_ssthresh = max(tp->snd_cwnd >> 1U, 2U);
	155	}
	156
	157	static void dctcp_state(struct sock *sk, u8 new_state)
	158	{
	159	if (new_state == TCP_CA_Recovery &&
	160	new_state != inet_csk(sk)->icsk_ca_state)
	161	dctcp_react_to_loss(sk);
	162	/* We handle RTO in dctcp_cwnd_event to ensure that we perform only
	163	* one loss-adjustment per RTT.
	164	*/
e3118e83 DB	165	}
e3118e83 DB	166
e3118e83 DB	167	static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
e3118e83 DB	168	{
ffd177de YC	169	struct dctcp *ca = inet_csk_ca(sk);
ffd177de YC	170
e3118e83 DB	171	switch (ev) {
e3118e83 DB	172	case CA_EVENT_ECN_IS_CE:
e3118e83	173	case CA_EVENT_ECN_NO_CE:
ffd177de	174	dctcp_ece_ack_update(sk, ev, &ca->prior_rcv_nxt, &ca->ce_state);
e3118e83	175	break;
aecfde23 KDS	176	case CA_EVENT_LOSS:
	177	dctcp_react_to_loss(sk);
	178	break;
e3118e83 DB	179	default:
	180	/* Don't care for the rest. */
	181	break;
	182	}
	183	}
	184
64f40ff5 ED	185	static size_t dctcp_get_info(struct sock sk, u32 ext, int attr,
64f40ff5 ED	186	union tcp_cc_info *info)
e3118e83 DB	187	{
e3118e83 DB	188	const struct dctcp *ca = inet_csk_ca(sk);
e3058450	189	const struct tcp_sock *tp = tcp_sk(sk);
e3118e83 DB	190
	191	/* Fill it also in case of VEGASINFO due to req struct limits.
	192	* We can still correctly retrieve it later.
	193	*/
	194	if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) \|\|
	195	ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
dcf1158b	196	memset(&info->dctcp, 0, sizeof(info->dctcp));
e3118e83	197	if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) {
64f40ff5 ED	198	info->dctcp.dctcp_enabled = 1;
	199	info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
	200	info->dctcp.dctcp_alpha = ca->dctcp_alpha;
e3058450 ED	201	info->dctcp.dctcp_ab_ecn = tp->mss_cache *
	202	(tp->delivered_ce - ca->old_delivered_ce);
	203	info->dctcp.dctcp_ab_tot = tp->mss_cache *
	204	(tp->delivered - ca->old_delivered);
e3118e83 DB	205	}
e3118e83 DB	206
64f40ff5	207	*attr = INET_DIAG_DCTCPINFO;
dcf1158b	208	return sizeof(info->dctcp);
e3118e83	209	}
521f1cf1	210	return 0;
e3118e83 DB	211	}
e3118e83 DB	212
ce6dd233 FW	213	static u32 dctcp_cwnd_undo(struct sock *sk)
	214	{
	215	const struct dctcp *ca = inet_csk_ca(sk);
	216
	217	return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
	218	}
	219
e3118e83 DB	220	static struct tcp_congestion_ops dctcp __read_mostly = {
	221	.init = dctcp_init,
	222	.in_ack_event = dctcp_update_alpha,
	223	.cwnd_event = dctcp_cwnd_event,
	224	.ssthresh = dctcp_ssthresh,
	225	.cong_avoid = tcp_reno_cong_avoid,
ce6dd233	226	.undo_cwnd = dctcp_cwnd_undo,
e3118e83 DB	227	.set_state = dctcp_state,
	228	.get_info = dctcp_get_info,
	229	.flags = TCP_CONG_NEEDS_ECN,
	230	.owner = THIS_MODULE,
	231	.name = "dctcp",
	232	};
	233
	234	static struct tcp_congestion_ops dctcp_reno __read_mostly = {
	235	.ssthresh = tcp_reno_ssthresh,
	236	.cong_avoid = tcp_reno_cong_avoid,
e9799183	237	.undo_cwnd = tcp_reno_undo_cwnd,
e3118e83 DB	238	.get_info = dctcp_get_info,
	239	.owner = THIS_MODULE,
	240	.name = "dctcp-reno",
	241	};
	242
	243	static int __init dctcp_register(void)
	244	{
	245	BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE);
	246	return tcp_register_congestion_control(&dctcp);
	247	}
	248
	249	static void __exit dctcp_unregister(void)
	250	{
	251	tcp_unregister_congestion_control(&dctcp);
	252	}
	253
	254	module_init(dctcp_register);
	255	module_exit(dctcp_unregister);
	256
	257	MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
	258	MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
	259	MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");
	260
	261	MODULE_LICENSE("GPL v2");
	262	MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");