linux_dsm_epyc7002/net/ipv4/tcp_lp.c
Yuchung Cheng 9f9843a751 tcp: properly handle stretch acks in slow start
Slow start now increases cwnd by 1 if an ACK acknowledges some packets,
regardless the number of packets. Consequently slow start performance
is highly dependent on the degree of the stretch ACKs caused by
receiver or network ACK compression mechanisms (e.g., delayed-ACK,
GRO, etc).  But slow start algorithm is to send twice the amount of
packets of packets left so it should process a stretch ACK of degree
N as if N ACKs of degree 1, then exits when cwnd exceeds ssthresh. A
follow up patch will use the remainder of the N (if greater than 1)
to adjust cwnd in the congestion avoidance phase.

In addition this patch retires the experimental limited slow start
(LSS) feature. LSS has multiple drawbacks but questionable benefit. The
fractional cwnd increase in LSS requires a loop in slow start even
though it's rarely used. Configuring such an increase step via a global
sysctl on different BDPS seems hard. Finally and most importantly the
slow start overshoot concern is now better covered by the Hybrid slow
start (hystart) enabled by default.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-11-04 19:57:59 -05:00

346 lines
8.8 KiB
C

/*
* TCP Low Priority (TCP-LP)
*
* TCP Low Priority is a distributed algorithm whose goal is to utilize only
* the excess network bandwidth as compared to the ``fair share`` of
* bandwidth as targeted by TCP.
*
* As of 2.6.13, Linux supports pluggable congestion control algorithms.
* Due to the limitation of the API, we take the following changes from
* the original TCP-LP implementation:
* o We use newReno in most core CA handling. Only add some checking
* within cong_avoid.
* o Error correcting in remote HZ, therefore remote HZ will be keeped
* on checking and updating.
* o Handling calculation of One-Way-Delay (OWD) within rtt_sample, since
* OWD have a similar meaning as RTT. Also correct the buggy formular.
* o Handle reaction for Early Congestion Indication (ECI) within
* pkts_acked, as mentioned within pseudo code.
* o OWD is handled in relative format, where local time stamp will in
* tcp_time_stamp format.
*
* Original Author:
* Aleksandar Kuzmanovic <akuzma@northwestern.edu>
* Available from:
* http://www.ece.rice.edu/~akuzma/Doc/akuzma/TCP-LP.pdf
* Original implementation for 2.4.19:
* http://www-ece.rice.edu/networks/TCP-LP/
*
* 2.6.x module Authors:
* Wong Hoi Sing, Edison <hswong3i@gmail.com>
* Hung Hing Lun, Mike <hlhung3i@gmail.com>
* SourceForge project page:
* http://tcp-lp-mod.sourceforge.net/
*/
#include <linux/module.h>
#include <net/tcp.h>
/* resolution of owd */
#define LP_RESOL 1000
/**
* enum tcp_lp_state
* @LP_VALID_RHZ: is remote HZ valid?
* @LP_VALID_OWD: is OWD valid?
* @LP_WITHIN_THR: are we within threshold?
* @LP_WITHIN_INF: are we within inference?
*
* TCP-LP's state flags.
* We create this set of state flag mainly for debugging.
*/
enum tcp_lp_state {
LP_VALID_RHZ = (1 << 0),
LP_VALID_OWD = (1 << 1),
LP_WITHIN_THR = (1 << 3),
LP_WITHIN_INF = (1 << 4),
};
/**
* struct lp
* @flag: TCP-LP state flag
* @sowd: smoothed OWD << 3
* @owd_min: min OWD
* @owd_max: max OWD
* @owd_max_rsv: resrved max owd
* @remote_hz: estimated remote HZ
* @remote_ref_time: remote reference time
* @local_ref_time: local reference time
* @last_drop: time for last active drop
* @inference: current inference
*
* TCP-LP's private struct.
* We get the idea from original TCP-LP implementation where only left those we
* found are really useful.
*/
struct lp {
u32 flag;
u32 sowd;
u32 owd_min;
u32 owd_max;
u32 owd_max_rsv;
u32 remote_hz;
u32 remote_ref_time;
u32 local_ref_time;
u32 last_drop;
u32 inference;
};
/**
* tcp_lp_init
*
* Init all required variables.
* Clone the handling from Vegas module implementation.
*/
static void tcp_lp_init(struct sock *sk)
{
struct lp *lp = inet_csk_ca(sk);
lp->flag = 0;
lp->sowd = 0;
lp->owd_min = 0xffffffff;
lp->owd_max = 0;
lp->owd_max_rsv = 0;
lp->remote_hz = 0;
lp->remote_ref_time = 0;
lp->local_ref_time = 0;
lp->last_drop = 0;
lp->inference = 0;
}
/**
* tcp_lp_cong_avoid
*
* Implementation of cong_avoid.
* Will only call newReno CA when away from inference.
* From TCP-LP's paper, this will be handled in additive increasement.
*/
static void tcp_lp_cong_avoid(struct sock *sk, u32 ack, u32 acked,
u32 in_flight)
{
struct lp *lp = inet_csk_ca(sk);
if (!(lp->flag & LP_WITHIN_INF))
tcp_reno_cong_avoid(sk, ack, acked, in_flight);
}
/**
* tcp_lp_remote_hz_estimator
*
* Estimate remote HZ.
* We keep on updating the estimated value, where original TCP-LP
* implementation only guest it for once and use forever.
*/
static u32 tcp_lp_remote_hz_estimator(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
struct lp *lp = inet_csk_ca(sk);
s64 rhz = lp->remote_hz << 6; /* remote HZ << 6 */
s64 m = 0;
/* not yet record reference time
* go away!! record it before come back!! */
if (lp->remote_ref_time == 0 || lp->local_ref_time == 0)
goto out;
/* we can't calc remote HZ with no different!! */
if (tp->rx_opt.rcv_tsval == lp->remote_ref_time ||
tp->rx_opt.rcv_tsecr == lp->local_ref_time)
goto out;
m = HZ * (tp->rx_opt.rcv_tsval -
lp->remote_ref_time) / (tp->rx_opt.rcv_tsecr -
lp->local_ref_time);
if (m < 0)
m = -m;
if (rhz > 0) {
m -= rhz >> 6; /* m is now error in remote HZ est */
rhz += m; /* 63/64 old + 1/64 new */
} else
rhz = m << 6;
out:
/* record time for successful remote HZ calc */
if ((rhz >> 6) > 0)
lp->flag |= LP_VALID_RHZ;
else
lp->flag &= ~LP_VALID_RHZ;
/* record reference time stamp */
lp->remote_ref_time = tp->rx_opt.rcv_tsval;
lp->local_ref_time = tp->rx_opt.rcv_tsecr;
return rhz >> 6;
}
/**
* tcp_lp_owd_calculator
*
* Calculate one way delay (in relative format).
* Original implement OWD as minus of remote time difference to local time
* difference directly. As this time difference just simply equal to RTT, when
* the network status is stable, remote RTT will equal to local RTT, and result
* OWD into zero.
* It seems to be a bug and so we fixed it.
*/
static u32 tcp_lp_owd_calculator(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
struct lp *lp = inet_csk_ca(sk);
s64 owd = 0;
lp->remote_hz = tcp_lp_remote_hz_estimator(sk);
if (lp->flag & LP_VALID_RHZ) {
owd =
tp->rx_opt.rcv_tsval * (LP_RESOL / lp->remote_hz) -
tp->rx_opt.rcv_tsecr * (LP_RESOL / HZ);
if (owd < 0)
owd = -owd;
}
if (owd > 0)
lp->flag |= LP_VALID_OWD;
else
lp->flag &= ~LP_VALID_OWD;
return owd;
}
/**
* tcp_lp_rtt_sample
*
* Implementation or rtt_sample.
* Will take the following action,
* 1. calc OWD,
* 2. record the min/max OWD,
* 3. calc smoothed OWD (SOWD).
* Most ideas come from the original TCP-LP implementation.
*/
static void tcp_lp_rtt_sample(struct sock *sk, u32 rtt)
{
struct lp *lp = inet_csk_ca(sk);
s64 mowd = tcp_lp_owd_calculator(sk);
/* sorry that we don't have valid data */
if (!(lp->flag & LP_VALID_RHZ) || !(lp->flag & LP_VALID_OWD))
return;
/* record the next min owd */
if (mowd < lp->owd_min)
lp->owd_min = mowd;
/* always forget the max of the max
* we just set owd_max as one below it */
if (mowd > lp->owd_max) {
if (mowd > lp->owd_max_rsv) {
if (lp->owd_max_rsv == 0)
lp->owd_max = mowd;
else
lp->owd_max = lp->owd_max_rsv;
lp->owd_max_rsv = mowd;
} else
lp->owd_max = mowd;
}
/* calc for smoothed owd */
if (lp->sowd != 0) {
mowd -= lp->sowd >> 3; /* m is now error in owd est */
lp->sowd += mowd; /* owd = 7/8 owd + 1/8 new */
} else
lp->sowd = mowd << 3; /* take the measured time be owd */
}
/**
* tcp_lp_pkts_acked
*
* Implementation of pkts_acked.
* Deal with active drop under Early Congestion Indication.
* Only drop to half and 1 will be handle, because we hope to use back
* newReno in increase case.
* We work it out by following the idea from TCP-LP's paper directly
*/
static void tcp_lp_pkts_acked(struct sock *sk, u32 num_acked, s32 rtt_us)
{
struct tcp_sock *tp = tcp_sk(sk);
struct lp *lp = inet_csk_ca(sk);
if (rtt_us > 0)
tcp_lp_rtt_sample(sk, rtt_us);
/* calc inference */
if (tcp_time_stamp > tp->rx_opt.rcv_tsecr)
lp->inference = 3 * (tcp_time_stamp - tp->rx_opt.rcv_tsecr);
/* test if within inference */
if (lp->last_drop && (tcp_time_stamp - lp->last_drop < lp->inference))
lp->flag |= LP_WITHIN_INF;
else
lp->flag &= ~LP_WITHIN_INF;
/* test if within threshold */
if (lp->sowd >> 3 <
lp->owd_min + 15 * (lp->owd_max - lp->owd_min) / 100)
lp->flag |= LP_WITHIN_THR;
else
lp->flag &= ~LP_WITHIN_THR;
pr_debug("TCP-LP: %05o|%5u|%5u|%15u|%15u|%15u\n", lp->flag,
tp->snd_cwnd, lp->remote_hz, lp->owd_min, lp->owd_max,
lp->sowd >> 3);
if (lp->flag & LP_WITHIN_THR)
return;
/* FIXME: try to reset owd_min and owd_max here
* so decrease the chance the min/max is no longer suitable
* and will usually within threshold when whithin inference */
lp->owd_min = lp->sowd >> 3;
lp->owd_max = lp->sowd >> 2;
lp->owd_max_rsv = lp->sowd >> 2;
/* happened within inference
* drop snd_cwnd into 1 */
if (lp->flag & LP_WITHIN_INF)
tp->snd_cwnd = 1U;
/* happened after inference
* cut snd_cwnd into half */
else
tp->snd_cwnd = max(tp->snd_cwnd >> 1U, 1U);
/* record this drop time */
lp->last_drop = tcp_time_stamp;
}
static struct tcp_congestion_ops tcp_lp __read_mostly = {
.flags = TCP_CONG_RTT_STAMP,
.init = tcp_lp_init,
.ssthresh = tcp_reno_ssthresh,
.cong_avoid = tcp_lp_cong_avoid,
.min_cwnd = tcp_reno_min_cwnd,
.pkts_acked = tcp_lp_pkts_acked,
.owner = THIS_MODULE,
.name = "lp"
};
static int __init tcp_lp_register(void)
{
BUILD_BUG_ON(sizeof(struct lp) > ICSK_CA_PRIV_SIZE);
return tcp_register_congestion_control(&tcp_lp);
}
static void __exit tcp_lp_unregister(void)
{
tcp_unregister_congestion_control(&tcp_lp);
}
module_init(tcp_lp_register);
module_exit(tcp_lp_unregister);
MODULE_AUTHOR("Wong Hoi Sing Edison, Hung Hing Lun Mike");
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("TCP Low Priority");