svn commit: r216114 - in head/sys: modules/cc modules/cc/cc_cubic
netinet/cc
Lawrence Stewart
lstewart at FreeBSD.org
Thu Dec 2 06:05:44 UTC 2010
Author: lstewart
Date: Thu Dec 2 06:05:44 2010
New Revision: 216114
URL: http://svn.freebsd.org/changeset/base/216114
Log:
Import a clean-room implementation of the experimental CUBIC congestion control
algorithm based on the Internet-Draft "draft-rhee-tcpm-cubic-02.txt". It is
implemented as a kernel module compatible with the recently committed modular
congestion control framework.
CUBIC was designed for provide increased throughput in fast and long-distance
networks. It attempts to maintain fairness when competing with legacy NewReno
TCP in lower speed scenarios where NewReno is able to operate adequately. The
paper "CUBIC: A New TCP-Friendly High-Speed TCP Variant" provides additional
detail.
In collaboration with: David Hayes <dahayes at swin edu au> and
Grenville Armitage <garmitage at swin edu au>
Sponsored by: FreeBSD Foundation
Reviewed by: rpaulo (older patch from a few weeks ago)
MFC after: 3 months
Added:
head/sys/modules/cc/cc_cubic/
head/sys/modules/cc/cc_cubic/Makefile (contents, props changed)
head/sys/netinet/cc/cc_cubic.c (contents, props changed)
head/sys/netinet/cc/cc_cubic.h (contents, props changed)
Modified:
head/sys/modules/cc/Makefile
Modified: head/sys/modules/cc/Makefile
==============================================================================
--- head/sys/modules/cc/Makefile Thu Dec 2 04:58:07 2010 (r216113)
+++ head/sys/modules/cc/Makefile Thu Dec 2 06:05:44 2010 (r216114)
@@ -1,5 +1,5 @@
# $FreeBSD$
-SUBDIR=
+SUBDIR= cc_cubic
.include <bsd.subdir.mk>
Added: head/sys/modules/cc/cc_cubic/Makefile
==============================================================================
--- /dev/null 00:00:00 1970 (empty, because file is newly added)
+++ head/sys/modules/cc/cc_cubic/Makefile Thu Dec 2 06:05:44 2010 (r216114)
@@ -0,0 +1,9 @@
+# $FreeBSD$
+
+.include <bsd.own.mk>
+
+.PATH: ${.CURDIR}/../../../netinet/cc
+KMOD= cc_cubic
+SRCS= cc_cubic.c
+
+.include <bsd.kmod.mk>
Added: head/sys/netinet/cc/cc_cubic.c
==============================================================================
--- /dev/null 00:00:00 1970 (empty, because file is newly added)
+++ head/sys/netinet/cc/cc_cubic.c Thu Dec 2 06:05:44 2010 (r216114)
@@ -0,0 +1,396 @@
+/*-
+ * Copyright (c) 2008-2010 Lawrence Stewart <lstewart at freebsd.org>
+ * Copyright (c) 2010 The FreeBSD Foundation
+ * All rights reserved.
+ *
+ * This software was developed by Lawrence Stewart while studying at the Centre
+ * for Advanced Internet Architectures, Swinburne University, 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 CUBIC congestion control algorithm for FreeBSD,
+ * based on the Internet Draft "draft-rhee-tcpm-cubic-02" by Rhee, Xu and Ha.
+ * Originally released as part of the NewTCP research project at Swinburne
+ * University'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/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/cc.h>
+#include <netinet/tcp_seq.h>
+#include <netinet/tcp_timer.h>
+#include <netinet/tcp_var.h>
+
+#include <netinet/cc/cc_cubic.h>
+#include <netinet/cc/cc_module.h>
+
+static void cubic_ack_received(struct cc_var *ccv, uint16_t type);
+static void cubic_cb_destroy(struct cc_var *ccv);
+static int cubic_cb_init(struct cc_var *ccv);
+static void cubic_cong_signal(struct cc_var *ccv, uint32_t type);
+static void cubic_conn_init(struct cc_var *ccv);
+static int cubic_mod_init(void);
+static void cubic_post_recovery(struct cc_var *ccv);
+static void cubic_record_rtt(struct cc_var *ccv);
+static void cubic_ssthresh_update(struct cc_var *ccv);
+
+struct cubic {
+ /* Cubic K in fixed point form with CUBIC_SHIFT worth of precision. */
+ int64_t K;
+ /* Sum of RTT samples across an epoch in ticks. */
+ int64_t sum_rtt_ticks;
+ /* cwnd at the most recent congestion event. */
+ unsigned long max_cwnd;
+ /* cwnd at the previous congestion event. */
+ unsigned long prev_max_cwnd;
+ /* Number of congestion events. */
+ uint32_t num_cong_events;
+ /* Minimum observed rtt in ticks. */
+ int min_rtt_ticks;
+ /* Mean observed rtt between congestion epochs. */
+ int mean_rtt_ticks;
+ /* ACKs since last congestion event. */
+ int epoch_ack_count;
+ /* Time of last congestion event in ticks. */
+ int t_last_cong;
+};
+
+MALLOC_DECLARE(M_CUBIC);
+MALLOC_DEFINE(M_CUBIC, "cubic data",
+ "Per connection data required for the CUBIC congestion control algorithm");
+
+struct cc_algo cubic_cc_algo = {
+ .name = "cubic",
+ .ack_received = cubic_ack_received,
+ .cb_destroy = cubic_cb_destroy,
+ .cb_init = cubic_cb_init,
+ .cong_signal = cubic_cong_signal,
+ .conn_init = cubic_conn_init,
+ .mod_init = cubic_mod_init,
+ .post_recovery = cubic_post_recovery,
+};
+
+static void
+cubic_ack_received(struct cc_var *ccv, uint16_t type)
+{
+ struct cubic *cubic_data;
+ unsigned long w_tf, w_cubic_next;
+ int ticks_since_cong;
+
+ cubic_data = ccv->cc_data;
+ cubic_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))) {
+ /* Use the logic in NewReno ack_received() for slow start. */
+ if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
+ cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)
+ newreno_cc_algo.ack_received(ccv, type);
+ else {
+ ticks_since_cong = ticks - cubic_data->t_last_cong;
+
+ /*
+ * The mean RTT is used to best reflect the equations in
+ * the I-D. Using min_rtt in the tf_cwnd calculation
+ * causes w_tf to grow much faster than it should if the
+ * RTT is dominated by network buffering rather than
+ * propogation delay.
+ */
+ w_tf = tf_cwnd(ticks_since_cong,
+ cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
+ CCV(ccv, t_maxseg));
+
+ w_cubic_next = cubic_cwnd(ticks_since_cong +
+ cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
+ CCV(ccv, t_maxseg), cubic_data->K);
+
+ ccv->flags &= ~CCF_ABC_SENTAWND;
+
+ if (w_cubic_next < w_tf)
+ /*
+ * TCP-friendly region, follow tf
+ * cwnd growth.
+ */
+ CCV(ccv, snd_cwnd) = w_tf;
+
+ else if (CCV(ccv, snd_cwnd) < w_cubic_next) {
+ /*
+ * Concave or convex region, follow CUBIC
+ * cwnd growth.
+ */
+ if (V_tcp_do_rfc3465)
+ CCV(ccv, snd_cwnd) = w_cubic_next;
+ else
+ CCV(ccv, snd_cwnd) += ((w_cubic_next -
+ CCV(ccv, snd_cwnd)) *
+ CCV(ccv, t_maxseg)) /
+ CCV(ccv, snd_cwnd);
+ }
+
+ /*
+ * If we're not in slow start and we're probing for a
+ * new cwnd limit at the start of a connection
+ * (happens when hostcache has a relevant entry),
+ * keep updating our current estimate of the
+ * max_cwnd.
+ */
+ if (cubic_data->num_cong_events == 0 &&
+ cubic_data->max_cwnd < CCV(ccv, snd_cwnd))
+ cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
+ }
+ }
+}
+
+static void
+cubic_cb_destroy(struct cc_var *ccv)
+{
+
+ if (ccv->cc_data != NULL)
+ free(ccv->cc_data, M_CUBIC);
+}
+
+static int
+cubic_cb_init(struct cc_var *ccv)
+{
+ struct cubic *cubic_data;
+
+ cubic_data = malloc(sizeof(struct cubic), M_CUBIC, M_NOWAIT|M_ZERO);
+
+ if (cubic_data == NULL)
+ return (ENOMEM);
+
+ /* Init some key variables with sensible defaults. */
+ cubic_data->t_last_cong = ticks;
+ cubic_data->min_rtt_ticks = TCPTV_SRTTBASE;
+ cubic_data->mean_rtt_ticks = TCPTV_SRTTBASE;
+
+ ccv->cc_data = cubic_data;
+
+ return (0);
+}
+
+/*
+ * Perform any necessary tasks before we enter congestion recovery.
+ */
+static void
+cubic_cong_signal(struct cc_var *ccv, uint32_t type)
+{
+ struct cubic *cubic_data;
+
+ cubic_data = ccv->cc_data;
+
+ switch (type) {
+ case CC_NDUPACK:
+ if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
+ if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
+ cubic_ssthresh_update(ccv);
+ cubic_data->num_cong_events++;
+ cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
+ cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
+ }
+ ENTER_RECOVERY(CCV(ccv, t_flags));
+ }
+ break;
+
+ case CC_ECN:
+ if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
+ cubic_ssthresh_update(ccv);
+ cubic_data->num_cong_events++;
+ cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
+ cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
+ cubic_data->t_last_cong = ticks;
+ CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
+ 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)
+ cubic_data->num_cong_events++;
+ cubic_data->t_last_cong = ticks;
+ break;
+ }
+}
+
+static void
+cubic_conn_init(struct cc_var *ccv)
+{
+ struct cubic *cubic_data;
+
+ cubic_data = ccv->cc_data;
+
+ /*
+ * Ensure we have a sane initial value for max_cwnd recorded. Without
+ * this here bad things happen when entries from the TCP hostcache
+ * get used.
+ */
+ cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
+}
+
+static int
+cubic_mod_init(void)
+{
+
+ cubic_cc_algo.after_idle = newreno_cc_algo.after_idle;
+
+ return (0);
+}
+
+/*
+ * Perform any necessary tasks before we exit congestion recovery.
+ */
+static void
+cubic_post_recovery(struct cc_var *ccv)
+{
+ struct cubic *cubic_data;
+
+ cubic_data = ccv->cc_data;
+
+ /* Fast convergence heuristic. */
+ if (cubic_data->max_cwnd < cubic_data->prev_max_cwnd)
+ cubic_data->max_cwnd = (cubic_data->max_cwnd * CUBIC_FC_FACTOR)
+ >> CUBIC_SHIFT;
+
+ 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 CUBIC method.
+ *
+ * XXXLAS: Find a way to do this without needing curack
+ */
+ if (SEQ_GT(ccv->curack + CCV(ccv, snd_ssthresh),
+ CCV(ccv, snd_max)))
+ CCV(ccv, snd_cwnd) = CCV(ccv, snd_max) - ccv->curack +
+ CCV(ccv, t_maxseg);
+ else
+ /* Update cwnd based on beta and adjusted max_cwnd. */
+ CCV(ccv, snd_cwnd) = max(1, ((CUBIC_BETA *
+ cubic_data->max_cwnd) >> CUBIC_SHIFT));
+ }
+ cubic_data->t_last_cong = ticks;
+
+ /* Calculate the average RTT between congestion epochs. */
+ if (cubic_data->epoch_ack_count > 0 && cubic_data->sum_rtt_ticks > 0)
+ cubic_data->mean_rtt_ticks = (int)(cubic_data->sum_rtt_ticks /
+ cubic_data->epoch_ack_count);
+ else
+ /* For safety. */
+ cubic_data->mean_rtt_ticks = cubic_data->min_rtt_ticks;
+
+ cubic_data->epoch_ack_count = 0;
+ cubic_data->sum_rtt_ticks = 0;
+ cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
+}
+
+/*
+ * Record the min RTT and sum samples for the epoch average RTT calculation.
+ */
+static void
+cubic_record_rtt(struct cc_var *ccv)
+{
+ struct cubic *cubic_data;
+ int t_srtt_ticks;
+
+ /* Ignore srtt until a min number of samples have been taken. */
+ if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) {
+ cubic_data = ccv->cc_data;
+ t_srtt_ticks = CCV(ccv, t_srtt) / TCP_RTT_SCALE;
+
+ /*
+ * Record the current SRTT as our minrtt if it's the smallest
+ * we've seen or minrtt is currently equal to its initialised
+ * value.
+ *
+ * XXXLAS: Should there be some hysteresis for minrtt?
+ */
+ if ((t_srtt_ticks < cubic_data->min_rtt_ticks ||
+ cubic_data->min_rtt_ticks == TCPTV_SRTTBASE))
+ cubic_data->min_rtt_ticks = max(1, t_srtt_ticks);
+
+ /* Sum samples for epoch average RTT calculation. */
+ cubic_data->sum_rtt_ticks += t_srtt_ticks;
+ cubic_data->epoch_ack_count++;
+ }
+}
+
+/*
+ * Update the ssthresh in the event of congestion.
+ */
+static void
+cubic_ssthresh_update(struct cc_var *ccv)
+{
+ struct cubic *cubic_data;
+
+ cubic_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 (cubic_data->num_cong_events == 0)
+ CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd) >> 1;
+ else
+ CCV(ccv, snd_ssthresh) = (CCV(ccv, snd_cwnd) * CUBIC_BETA)
+ >> CUBIC_SHIFT;
+}
+
+
+DECLARE_CC_MODULE(cubic, &cubic_cc_algo);
Added: head/sys/netinet/cc/cc_cubic.h
==============================================================================
--- /dev/null 00:00:00 1970 (empty, because file is newly added)
+++ head/sys/netinet/cc/cc_cubic.h Thu Dec 2 06:05:44 2010 (r216114)
@@ -0,0 +1,229 @@
+/*-
+ * Copyright (c) 2008-2010 Lawrence Stewart <lstewart at freebsd.org>
+ * Copyright (c) 2010 The FreeBSD Foundation
+ * All rights reserved.
+ *
+ * This software was developed by Lawrence Stewart while studying at the Centre
+ * for Advanced Internet Architectures, Swinburne University, 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.
+ *
+ * $FreeBSD$
+ */
+
+#ifndef _NETINET_CC_CUBIC_H_
+#define _NETINET_CC_CUBIC_H_
+
+/* Number of bits of precision for fixed point math calcs. */
+#define CUBIC_SHIFT 8
+
+#define CUBIC_SHIFT_4 32
+
+/* 0.5 << CUBIC_SHIFT. */
+#define RENO_BETA 128
+
+/* ~0.8 << CUBIC_SHIFT. */
+#define CUBIC_BETA 204
+
+/* ~0.2 << CUBIC_SHIFT. */
+#define ONE_SUB_CUBIC_BETA 51
+
+/* 3 * ONE_SUB_CUBIC_BETA. */
+#define THREE_X_PT2 153
+
+/* (2 << CUBIC_SHIFT) - ONE_SUB_CUBIC_BETA. */
+#define TWO_SUB_PT2 461
+
+/* ~0.4 << CUBIC_SHIFT. */
+#define CUBIC_C_FACTOR 102
+
+/* CUBIC fast convergence factor: ~0.9 << CUBIC_SHIFT. */
+#define CUBIC_FC_FACTOR 230
+
+/* Don't trust s_rtt until this many rtt samples have been taken. */
+#define CUBIC_MIN_RTT_SAMPLES 8
+
+/* Userland only bits. */
+#ifndef _KERNEL
+
+extern int hz;
+
+/*
+ * Implementation based on the formulae found in the CUBIC Internet Draft
+ * "draft-rhee-tcpm-cubic-02".
+ *
+ * Note BETA used in cc_cubic is equal to (1-beta) in the I-D
+ */
+
+static __inline float
+theoretical_cubic_k(double wmax_pkts)
+{
+ double C;
+
+ C = 0.4;
+
+ return (pow((wmax_pkts * 0.2) / C, (1.0 / 3.0)) * pow(2, CUBIC_SHIFT));
+}
+
+static __inline unsigned long
+theoretical_cubic_cwnd(int ticks_since_cong, unsigned long wmax, uint32_t smss)
+{
+ double C, wmax_pkts;
+
+ C = 0.4;
+ wmax_pkts = wmax / (double)smss;
+
+ return (smss * (wmax_pkts +
+ (C * pow(ticks_since_cong / (double)hz -
+ theoretical_cubic_k(wmax_pkts) / pow(2, CUBIC_SHIFT), 3.0))));
+}
+
+static __inline unsigned long
+theoretical_reno_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax,
+ uint32_t smss)
+{
+
+ return ((wmax * 0.5) + ((ticks_since_cong / (float)rtt_ticks) * smss));
+}
+
+static __inline unsigned long
+theoretical_tf_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax,
+ uint32_t smss)
+{
+
+ return ((wmax * 0.8) + ((3 * 0.2) / (2 - 0.2) *
+ (ticks_since_cong / (float)rtt_ticks) * smss));
+}
+
+#endif /* !_KERNEL */
+
+/*
+ * Compute the CUBIC K value used in the cwnd calculation, using an
+ * implementation of eqn 2 in the I-D. The method used
+ * here is adapted from Apple Computer Technical Report #KT-32.
+ */
+static __inline int64_t
+cubic_k(unsigned long wmax_pkts)
+{
+ int64_t s, K;
+ uint16_t p;
+
+ K = s = 0;
+ p = 0;
+
+ /* (wmax * beta)/C with CUBIC_SHIFT worth of precision. */
+ s = ((wmax_pkts * ONE_SUB_CUBIC_BETA) << CUBIC_SHIFT) / CUBIC_C_FACTOR;
+
+ /* Rebase s to be between 1 and 1/8 with a shift of CUBIC_SHIFT. */
+ while (s >= 256) {
+ s >>= 3;
+ p++;
+ }
+
+ /*
+ * Some magic constants taken from the Apple TR with appropriate
+ * shifts: 275 == 1.072302 << CUBIC_SHIFT, 98 == 0.3812513 <<
+ * CUBIC_SHIFT, 120 == 0.46946116 << CUBIC_SHIFT.
+ */
+ K = (((s * 275) >> CUBIC_SHIFT) + 98) -
+ (((s * s * 120) >> CUBIC_SHIFT) >> CUBIC_SHIFT);
+
+ /* Multiply by 2^p to undo the rebasing of s from above. */
+ return (K <<= p);
+}
+
+/*
+ * Compute the new cwnd value using an implementation of eqn 1 from the I-D.
+ * Thanks to Kip Macy for help debugging this function.
+ *
+ * XXXLAS: Characterise bounds for overflow.
+ */
+static __inline unsigned long
+cubic_cwnd(int ticks_since_cong, unsigned long wmax, uint32_t smss, int64_t K)
+{
+ int64_t cwnd;
+
+ /* K is in fixed point form with CUBIC_SHIFT worth of precision. */
+
+ /* t - K, with CUBIC_SHIFT worth of precision. */
+ cwnd = ((int64_t)(ticks_since_cong << CUBIC_SHIFT) - (K * hz)) / hz;
+
+ /* (t - K)^3, with CUBIC_SHIFT^3 worth of precision. */
+ cwnd *= (cwnd * cwnd);
+
+ /*
+ * C(t - K)^3 + wmax
+ * The down shift by CUBIC_SHIFT_4 is because cwnd has 4 lots of
+ * CUBIC_SHIFT included in the value. 3 from the cubing of cwnd above,
+ * and an extra from multiplying through by CUBIC_C_FACTOR.
+ */
+ cwnd = ((cwnd * CUBIC_C_FACTOR * smss) >> CUBIC_SHIFT_4) + wmax;
+
+ return ((unsigned long)cwnd);
+}
+
+/*
+ * Compute an approximation of the NewReno cwnd some number of ticks after a
+ * congestion event. RTT should be the average RTT estimate for the path
+ * measured over the previous congestion epoch and wmax is the value of cwnd at
+ * the last congestion event. The "TCP friendly" concept in the CUBIC I-D is
+ * rather tricky to understand and it turns out this function is not required.
+ * It is left here for reference.
+ */
+static __inline unsigned long
+reno_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax,
+ uint32_t smss)
+{
+
+ /*
+ * For NewReno, beta = 0.5, therefore: W_tcp(t) = wmax*0.5 + t/RTT
+ * W_tcp(t) deals with cwnd/wmax in pkts, so because our cwnd is in
+ * bytes, we have to multiply by smss.
+ */
+ return (((wmax * RENO_BETA) + (((ticks_since_cong * smss)
+ << CUBIC_SHIFT) / rtt_ticks)) >> CUBIC_SHIFT);
+}
+
+/*
+ * Compute an approximation of the "TCP friendly" cwnd some number of ticks
+ * after a congestion event that is designed to yield the same average cwnd as
+ * NewReno while using CUBIC's beta of 0.8. RTT should be the average RTT
+ * estimate for the path measured over the previous congestion epoch and wmax is
+ * the value of cwnd at the last congestion event.
+ */
+static __inline unsigned long
+tf_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax,
+ uint32_t smss)
+{
+
+ /* Equation 4 of I-D. */
+ return (((wmax * CUBIC_BETA) + (((THREE_X_PT2 * ticks_since_cong *
+ smss) << CUBIC_SHIFT) / TWO_SUB_PT2 / rtt_ticks)) >> CUBIC_SHIFT);
+}
+
+#endif /* _NETINET_CC_CUBIC_H_ */
More information about the svn-src-all
mailing list