freebsd-skq/sys/netinet/cc/cc_htcp.c
Hiren Panchasara e04310d59b Set slow start threshold more accurately on loss to be flightsize/2 instead of
cwnd/2 as recommended by RFC5681. (spotted by mmacy at nextbsd dot org)

Restore pre-r307901 behavior of aligning ssthresh/cwnd on mss boundary. (spotted
by slawa at zxy dot spb dot ru)

Tested by:	    dim, Slawa <slawa at zxy dot spb dot ru>
MFC after:	    1 month
X-MFC with:	    r307901
Sponsored by:	    Limelight Networks
Differential Revision:	https://reviews.freebsd.org/D8349
2016-11-01 21:08:37 +00:00

541 lines
17 KiB
C

/*-
* Copyright (c) 2007-2008
* Swinburne University of Technology, Melbourne, Australia
* Copyright (c) 2009-2010 Lawrence Stewart <lstewart@freebsd.org>
* Copyright (c) 2010 The FreeBSD Foundation
* All rights reserved.
*
* This software was developed at the Centre for Advanced Internet
* Architectures, Swinburne University of Technology, by Lawrence Stewart and
* James Healy, made possible in part by a grant from the Cisco University
* Research Program Fund at Community Foundation Silicon Valley.
*
* Portions of this software were developed at the Centre for Advanced
* Internet Architectures, Swinburne University of Technology, Melbourne,
* Australia by David Hayes under sponsorship from the FreeBSD Foundation.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
/*
* An implementation of the H-TCP congestion control algorithm for FreeBSD,
* based on the Internet Draft "draft-leith-tcp-htcp-06.txt" by Leith and
* Shorten. Originally released as part of the NewTCP research project at
* Swinburne University of Technology's Centre for Advanced Internet
* Architectures, Melbourne, Australia, which was made possible in part by a
* grant from the Cisco University Research Program Fund at Community Foundation
* Silicon Valley. More details are available at:
* http://caia.swin.edu.au/urp/newtcp/
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include <sys/param.h>
#include <sys/kernel.h>
#include <sys/limits.h>
#include <sys/malloc.h>
#include <sys/module.h>
#include <sys/socket.h>
#include <sys/socketvar.h>
#include <sys/sysctl.h>
#include <sys/systm.h>
#include <net/vnet.h>
#include <netinet/tcp.h>
#include <netinet/tcp_seq.h>
#include <netinet/tcp_timer.h>
#include <netinet/tcp_var.h>
#include <netinet/cc/cc.h>
#include <netinet/cc/cc_module.h>
/* Fixed point math shifts. */
#define HTCP_SHIFT 8
#define HTCP_ALPHA_INC_SHIFT 4
#define HTCP_INIT_ALPHA 1
#define HTCP_DELTA_L hz /* 1 sec in ticks. */
#define HTCP_MINBETA 128 /* 0.5 << HTCP_SHIFT. */
#define HTCP_MAXBETA 204 /* ~0.8 << HTCP_SHIFT. */
#define HTCP_MINROWE 26 /* ~0.1 << HTCP_SHIFT. */
#define HTCP_MAXROWE 512 /* 2 << HTCP_SHIFT. */
/* RTT_ref (ms) used in the calculation of alpha if RTT scaling is enabled. */
#define HTCP_RTT_REF 100
/* Don't trust SRTT until this many samples have been taken. */
#define HTCP_MIN_RTT_SAMPLES 8
/*
* HTCP_CALC_ALPHA performs a fixed point math calculation to determine the
* value of alpha, based on the function defined in the HTCP spec.
*
* i.e. 1 + 10(delta - delta_l) + ((delta - delta_l) / 2) ^ 2
*
* "diff" is passed in to the macro as "delta - delta_l" and is expected to be
* in units of ticks.
*
* The joyousnous of fixed point maths means our function implementation looks a
* little funky...
*
* In order to maintain some precision in the calculations, a fixed point shift
* HTCP_ALPHA_INC_SHIFT is used to ensure the integer divisions don't
* truncate the results too badly.
*
* The "16" value is the "1" term in the alpha function shifted up by
* HTCP_ALPHA_INC_SHIFT
*
* The "160" value is the "10" multiplier in the alpha function multiplied by
* 2^HTCP_ALPHA_INC_SHIFT
*
* Specifying these as constants reduces the computations required. After
* up-shifting all the terms in the function and performing the required
* calculations, we down-shift the final result by HTCP_ALPHA_INC_SHIFT to
* ensure it is back in the correct range.
*
* The "hz" terms are required as kernels can be configured to run with
* different tick timers, which we have to adjust for in the alpha calculation
* (which originally was defined in terms of seconds).
*
* We also have to be careful to constrain the value of diff such that it won't
* overflow whilst performing the calculation. The middle term i.e. (160 * diff)
* / hz is the limiting factor in the calculation. We must constrain diff to be
* less than the max size of an int divided by the constant 160 figure
* i.e. diff < INT_MAX / 160
*
* NB: Changing HTCP_ALPHA_INC_SHIFT will require you to MANUALLY update the
* constants used in this function!
*/
#define HTCP_CALC_ALPHA(diff) \
((\
(16) + \
((160 * (diff)) / hz) + \
(((diff) / hz) * (((diff) << HTCP_ALPHA_INC_SHIFT) / (4 * hz))) \
) >> HTCP_ALPHA_INC_SHIFT)
static void htcp_ack_received(struct cc_var *ccv, uint16_t type);
static void htcp_cb_destroy(struct cc_var *ccv);
static int htcp_cb_init(struct cc_var *ccv);
static void htcp_cong_signal(struct cc_var *ccv, uint32_t type);
static int htcp_mod_init(void);
static void htcp_post_recovery(struct cc_var *ccv);
static void htcp_recalc_alpha(struct cc_var *ccv);
static void htcp_recalc_beta(struct cc_var *ccv);
static void htcp_record_rtt(struct cc_var *ccv);
static void htcp_ssthresh_update(struct cc_var *ccv);
struct htcp {
/* cwnd before entering cong recovery. */
unsigned long prev_cwnd;
/* cwnd additive increase parameter. */
int alpha;
/* cwnd multiplicative decrease parameter. */
int beta;
/* Largest rtt seen for the flow. */
int maxrtt;
/* Shortest rtt seen for the flow. */
int minrtt;
/* Time of last congestion event in ticks. */
int t_last_cong;
};
static int htcp_rtt_ref;
/*
* The maximum number of ticks the value of diff can reach in
* htcp_recalc_alpha() before alpha will stop increasing due to overflow.
* See comment above HTCP_CALC_ALPHA for more info.
*/
static int htcp_max_diff = INT_MAX / ((1 << HTCP_ALPHA_INC_SHIFT) * 10);
/* Per-netstack vars. */
static VNET_DEFINE(u_int, htcp_adaptive_backoff) = 0;
static VNET_DEFINE(u_int, htcp_rtt_scaling) = 0;
#define V_htcp_adaptive_backoff VNET(htcp_adaptive_backoff)
#define V_htcp_rtt_scaling VNET(htcp_rtt_scaling)
static MALLOC_DEFINE(M_HTCP, "htcp data",
"Per connection data required for the HTCP congestion control algorithm");
struct cc_algo htcp_cc_algo = {
.name = "htcp",
.ack_received = htcp_ack_received,
.cb_destroy = htcp_cb_destroy,
.cb_init = htcp_cb_init,
.cong_signal = htcp_cong_signal,
.mod_init = htcp_mod_init,
.post_recovery = htcp_post_recovery,
};
static void
htcp_ack_received(struct cc_var *ccv, uint16_t type)
{
struct htcp *htcp_data;
htcp_data = ccv->cc_data;
htcp_record_rtt(ccv);
/*
* Regular ACK and we're not in cong/fast recovery and we're cwnd
* limited and we're either not doing ABC or are slow starting or are
* doing ABC and we've sent a cwnd's worth of bytes.
*/
if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
(ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 ||
CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
(V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) {
htcp_recalc_beta(ccv);
htcp_recalc_alpha(ccv);
/*
* Use the logic in NewReno ack_received() for slow start and
* for the first HTCP_DELTA_L ticks after either the flow starts
* or a congestion event (when alpha equals 1).
*/
if (htcp_data->alpha == 1 ||
CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh))
newreno_cc_algo.ack_received(ccv, type);
else {
if (V_tcp_do_rfc3465) {
/* Increment cwnd by alpha segments. */
CCV(ccv, snd_cwnd) += htcp_data->alpha *
CCV(ccv, t_maxseg);
ccv->flags &= ~CCF_ABC_SENTAWND;
} else
/*
* Increment cwnd by alpha/cwnd segments to
* approximate an increase of alpha segments
* per RTT.
*/
CCV(ccv, snd_cwnd) += (((htcp_data->alpha <<
HTCP_SHIFT) / (CCV(ccv, snd_cwnd) /
CCV(ccv, t_maxseg))) * CCV(ccv, t_maxseg))
>> HTCP_SHIFT;
}
}
}
static void
htcp_cb_destroy(struct cc_var *ccv)
{
if (ccv->cc_data != NULL)
free(ccv->cc_data, M_HTCP);
}
static int
htcp_cb_init(struct cc_var *ccv)
{
struct htcp *htcp_data;
htcp_data = malloc(sizeof(struct htcp), M_HTCP, M_NOWAIT);
if (htcp_data == NULL)
return (ENOMEM);
/* Init some key variables with sensible defaults. */
htcp_data->alpha = HTCP_INIT_ALPHA;
htcp_data->beta = HTCP_MINBETA;
htcp_data->maxrtt = TCPTV_SRTTBASE;
htcp_data->minrtt = TCPTV_SRTTBASE;
htcp_data->prev_cwnd = 0;
htcp_data->t_last_cong = ticks;
ccv->cc_data = htcp_data;
return (0);
}
/*
* Perform any necessary tasks before we enter congestion recovery.
*/
static void
htcp_cong_signal(struct cc_var *ccv, uint32_t type)
{
struct htcp *htcp_data;
uint32_t cwin;
u_int mss;
htcp_data = ccv->cc_data;
cwin = CCV(ccv, snd_cwnd);
mss = CCV(ccv, t_maxseg);
switch (type) {
case CC_NDUPACK:
if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
/*
* Apply hysteresis to maxrtt to ensure
* reductions in the RTT are reflected in our
* measurements.
*/
htcp_data->maxrtt = (htcp_data->minrtt +
(htcp_data->maxrtt - htcp_data->minrtt) *
95) / 100;
htcp_ssthresh_update(ccv);
CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
htcp_data->t_last_cong = ticks;
htcp_data->prev_cwnd = cwin;
}
ENTER_RECOVERY(CCV(ccv, t_flags));
}
break;
case CC_ECN:
if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
/*
* Apply hysteresis to maxrtt to ensure reductions in
* the RTT are reflected in our measurements.
*/
htcp_data->maxrtt = (htcp_data->minrtt + (htcp_data->maxrtt -
htcp_data->minrtt) * 95) / 100;
htcp_ssthresh_update(ccv);
CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
htcp_data->t_last_cong = ticks;
htcp_data->prev_cwnd = cwin;
ENTER_CONGRECOVERY(CCV(ccv, t_flags));
}
break;
case CC_RTO:
/*
* Grab the current time and record it so we know when the
* most recent congestion event was. Only record it when the
* timeout has fired more than once, as there is a reasonable
* chance the first one is a false alarm and may not indicate
* congestion.
*/
if (CCV(ccv, t_rxtshift) >= 2)
htcp_data->t_last_cong = ticks;
CCV(ccv, snd_ssthresh) =
max((CCV(ccv, snd_max) - CCV(ccv, snd_una)) / 2 / mss, 2)
* mss;
CCV(ccv, snd_cwnd) = mss;
break;
}
}
static int
htcp_mod_init(void)
{
htcp_cc_algo.after_idle = newreno_cc_algo.after_idle;
/*
* HTCP_RTT_REF is defined in ms, and t_srtt in the tcpcb is stored in
* units of TCP_RTT_SCALE*hz. Scale HTCP_RTT_REF to be in the same units
* as t_srtt.
*/
htcp_rtt_ref = (HTCP_RTT_REF * TCP_RTT_SCALE * hz) / 1000;
return (0);
}
/*
* Perform any necessary tasks before we exit congestion recovery.
*/
static void
htcp_post_recovery(struct cc_var *ccv)
{
int pipe;
struct htcp *htcp_data;
pipe = 0;
htcp_data = ccv->cc_data;
if (IN_FASTRECOVERY(CCV(ccv, t_flags))) {
/*
* If inflight data is less than ssthresh, set cwnd
* conservatively to avoid a burst of data, as suggested in the
* NewReno RFC. Otherwise, use the HTCP method.
*
* XXXLAS: Find a way to do this without needing curack
*/
if (V_tcp_do_rfc6675_pipe)
pipe = tcp_compute_pipe(ccv->ccvc.tcp);
else
pipe = CCV(ccv, snd_max) - ccv->curack;
if (pipe < CCV(ccv, snd_ssthresh))
CCV(ccv, snd_cwnd) = pipe + CCV(ccv, t_maxseg);
else
CCV(ccv, snd_cwnd) = max(1, ((htcp_data->beta *
htcp_data->prev_cwnd / CCV(ccv, t_maxseg))
>> HTCP_SHIFT)) * CCV(ccv, t_maxseg);
}
}
static void
htcp_recalc_alpha(struct cc_var *ccv)
{
struct htcp *htcp_data;
int alpha, diff, now;
htcp_data = ccv->cc_data;
now = ticks;
/*
* If ticks has wrapped around (will happen approximately once every 49
* days on a machine with the default kern.hz=1000) and a flow straddles
* the wrap point, our alpha calcs will be completely wrong. We cut our
* losses and restart alpha from scratch by setting t_last_cong = now -
* HTCP_DELTA_L.
*
* This does not deflate our cwnd at all. It simply slows the rate cwnd
* is growing by until alpha regains the value it held prior to taking
* this drastic measure.
*/
if (now < htcp_data->t_last_cong)
htcp_data->t_last_cong = now - HTCP_DELTA_L;
diff = now - htcp_data->t_last_cong - HTCP_DELTA_L;
/* Cap alpha if the value of diff would overflow HTCP_CALC_ALPHA(). */
if (diff < htcp_max_diff) {
/*
* If it has been more than HTCP_DELTA_L ticks since congestion,
* increase alpha according to the function defined in the spec.
*/
if (diff > 0) {
alpha = HTCP_CALC_ALPHA(diff);
/*
* Adaptive backoff fairness adjustment:
* 2 * (1 - beta) * alpha_raw
*/
if (V_htcp_adaptive_backoff)
alpha = max(1, (2 * ((1 << HTCP_SHIFT) -
htcp_data->beta) * alpha) >> HTCP_SHIFT);
/*
* RTT scaling: (RTT / RTT_ref) * alpha
* alpha will be the raw value from HTCP_CALC_ALPHA() if
* adaptive backoff is off, or the adjusted value if
* adaptive backoff is on.
*/
if (V_htcp_rtt_scaling)
alpha = max(1, (min(max(HTCP_MINROWE,
(CCV(ccv, t_srtt) << HTCP_SHIFT) /
htcp_rtt_ref), HTCP_MAXROWE) * alpha)
>> HTCP_SHIFT);
} else
alpha = 1;
htcp_data->alpha = alpha;
}
}
static void
htcp_recalc_beta(struct cc_var *ccv)
{
struct htcp *htcp_data;
htcp_data = ccv->cc_data;
/*
* TCPTV_SRTTBASE is the initialised value of each connection's SRTT, so
* we only calc beta if the connection's SRTT has been changed from its
* initial value. beta is bounded to ensure it is always between
* HTCP_MINBETA and HTCP_MAXBETA.
*/
if (V_htcp_adaptive_backoff && htcp_data->minrtt != TCPTV_SRTTBASE &&
htcp_data->maxrtt != TCPTV_SRTTBASE)
htcp_data->beta = min(max(HTCP_MINBETA,
(htcp_data->minrtt << HTCP_SHIFT) / htcp_data->maxrtt),
HTCP_MAXBETA);
else
htcp_data->beta = HTCP_MINBETA;
}
/*
* Record the minimum and maximum RTT seen for the connection. These are used in
* the calculation of beta if adaptive backoff is enabled.
*/
static void
htcp_record_rtt(struct cc_var *ccv)
{
struct htcp *htcp_data;
htcp_data = ccv->cc_data;
/* XXXLAS: Should there be some hysteresis for minrtt? */
/*
* Record the current SRTT as our minrtt if it's the smallest we've seen
* or minrtt is currently equal to its initialised value. Ignore SRTT
* until a min number of samples have been taken.
*/
if ((CCV(ccv, t_srtt) < htcp_data->minrtt ||
htcp_data->minrtt == TCPTV_SRTTBASE) &&
(CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES))
htcp_data->minrtt = CCV(ccv, t_srtt);
/*
* Record the current SRTT as our maxrtt if it's the largest we've
* seen. Ignore SRTT until a min number of samples have been taken.
*/
if (CCV(ccv, t_srtt) > htcp_data->maxrtt
&& CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES)
htcp_data->maxrtt = CCV(ccv, t_srtt);
}
/*
* Update the ssthresh in the event of congestion.
*/
static void
htcp_ssthresh_update(struct cc_var *ccv)
{
struct htcp *htcp_data;
htcp_data = ccv->cc_data;
/*
* On the first congestion event, set ssthresh to cwnd * 0.5, on
* subsequent congestion events, set it to cwnd * beta.
*/
if (CCV(ccv, snd_ssthresh) == TCP_MAXWIN << TCP_MAX_WINSHIFT)
CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) *
HTCP_MINBETA) >> HTCP_SHIFT;
else {
htcp_recalc_beta(ccv);
CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) *
htcp_data->beta) >> HTCP_SHIFT;
}
/* Align ssthresh to MSS boundary */
CCV(ccv, snd_ssthresh) = (CCV(ccv, snd_ssthresh) / CCV(ccv, t_maxseg))
* CCV(ccv, t_maxseg);
}
SYSCTL_DECL(_net_inet_tcp_cc_htcp);
SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, htcp, CTLFLAG_RW,
NULL, "H-TCP related settings");
SYSCTL_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, adaptive_backoff,
CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(htcp_adaptive_backoff), 0,
"enable H-TCP adaptive backoff");
SYSCTL_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, rtt_scaling,
CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(htcp_rtt_scaling), 0,
"enable H-TCP RTT scaling");
DECLARE_CC_MODULE(htcp, &htcp_cc_algo);