diff options
Diffstat (limited to 'sys/netinet')
| -rw-r--r-- | sys/netinet/cc/cc_cubic.c | 396 | ||||
| -rw-r--r-- | sys/netinet/cc/cc_cubic.h | 229 | ||||
| -rw-r--r-- | sys/netinet/cc/cc_htcp.c | 521 | ||||
| -rw-r--r-- | sys/netinet/cc/cc_newreno.c | 79 | ||||
| -rw-r--r-- | sys/netinet/if_ether.c | 10 | ||||
| -rw-r--r-- | sys/netinet/in.c | 2 | ||||
| -rw-r--r-- | sys/netinet/ip6.h | 2 | ||||
| -rw-r--r-- | sys/netinet/tcp_input.c | 2 | ||||
| -rw-r--r-- | sys/netinet/tcp_output.c | 35 | ||||
| -rw-r--r-- | sys/netinet/tcp_timer.c | 2 |
10 files changed, 1199 insertions, 79 deletions
diff --git a/sys/netinet/cc/cc_cubic.c b/sys/netinet/cc/cc_cubic.c new file mode 100644 index 0000000..d546ee1 --- /dev/null +++ b/sys/netinet/cc/cc_cubic.c @@ -0,0 +1,396 @@ +/*- + * Copyright (c) 2008-2010 Lawrence Stewart <lstewart@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); diff --git a/sys/netinet/cc/cc_cubic.h b/sys/netinet/cc/cc_cubic.h new file mode 100644 index 0000000..cf3470b --- /dev/null +++ b/sys/netinet/cc/cc_cubic.h @@ -0,0 +1,229 @@ +/*- + * Copyright (c) 2008-2010 Lawrence Stewart <lstewart@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_ */ diff --git a/sys/netinet/cc/cc_htcp.c b/sys/netinet/cc/cc_htcp.c new file mode 100644 index 0000000..61175ac --- /dev/null +++ b/sys/netinet/cc/cc_htcp.c @@ -0,0 +1,521 @@ +/*- + * 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, 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'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/cc.h> +#include <netinet/tcp_seq.h> +#include <netinet/tcp_timer.h> +#include <netinet/tcp_var.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(uint8_t, htcp_adaptive_backoff) = 0; +static VNET_DEFINE(uint8_t, htcp_rtt_scaling) = 0; +#define V_htcp_adaptive_backoff VNET(htcp_adaptive_backoff) +#define V_htcp_rtt_scaling VNET(htcp_rtt_scaling) + +MALLOC_DECLARE(M_HTCP); +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; + + htcp_data = ccv->cc_data; + + 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); + htcp_data->t_last_cong = ticks; + htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); + } + 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 = CCV(ccv, snd_cwnd); + 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; + 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) +{ + struct htcp *htcp_data; + + 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 (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 + 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 + * inital 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) = (CCV(ccv, snd_cwnd) * HTCP_MINBETA) + >> HTCP_SHIFT; + else { + htcp_recalc_beta(ccv); + CCV(ccv, snd_ssthresh) = (CCV(ccv, snd_cwnd) * htcp_data->beta) + >> HTCP_SHIFT; + } +} + + +SYSCTL_DECL(_net_inet_tcp_cc_htcp); +SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, htcp, CTLFLAG_RW, + NULL, "H-TCP related settings"); +SYSCTL_VNET_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, adaptive_backoff, CTLFLAG_RW, + &VNET_NAME(htcp_adaptive_backoff), 0, "enable H-TCP adaptive backoff"); +SYSCTL_VNET_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, rtt_scaling, CTLFLAG_RW, + &VNET_NAME(htcp_rtt_scaling), 0, "enable H-TCP RTT scaling"); + +DECLARE_CC_MODULE(htcp, &htcp_cc_algo); diff --git a/sys/netinet/cc/cc_newreno.c b/sys/netinet/cc/cc_newreno.c index e383510..c095540 100644 --- a/sys/netinet/cc/cc_newreno.c +++ b/sys/netinet/cc/cc_newreno.c @@ -52,41 +52,35 @@ __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/if.h> -#include <net/if_var.h> +#include <net/vnet.h> #include <netinet/cc.h> -#include <netinet/in.h> -#include <netinet/in_pcb.h> #include <netinet/tcp_seq.h> #include <netinet/tcp_var.h> #include <netinet/cc/cc_module.h> -void newreno_ack_received(struct cc_var *ccv, uint16_t type); -void newreno_cong_signal(struct cc_var *ccv, uint32_t type); -void newreno_post_recovery(struct cc_var *ccv); -void newreno_after_idle(struct cc_var *ccv); +static void newreno_ack_received(struct cc_var *ccv, uint16_t type); +static void newreno_after_idle(struct cc_var *ccv); +static void newreno_cong_signal(struct cc_var *ccv, uint32_t type); +static void newreno_post_recovery(struct cc_var *ccv); struct cc_algo newreno_cc_algo = { .name = "newreno", .ack_received = newreno_ack_received, + .after_idle = newreno_after_idle, .cong_signal = newreno_cong_signal, .post_recovery = newreno_post_recovery, - .after_idle = newreno_after_idle }; -/* - * Increase cwnd on receipt of a successful ACK: - * if cwnd <= ssthresh, increases by 1 MSS per ACK - * if cwnd > ssthresh, increase by ~1 MSS per RTT - */ -void +static void newreno_ack_received(struct cc_var *ccv, uint16_t type) { if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && @@ -153,10 +147,37 @@ newreno_ack_received(struct cc_var *ccv, uint16_t type) } } +static void +newreno_after_idle(struct cc_var *ccv) +{ + int rw; + + /* + * If we've been idle for more than one retransmit timeout the old + * congestion window is no longer current and we have to reduce it to + * the restart window before we can transmit again. + * + * The restart window is the initial window or the last CWND, whichever + * is smaller. + * + * This is done to prevent us from flooding the path with a full CWND at + * wirespeed, overloading router and switch buffers along the way. + * + * See RFC5681 Section 4.1. "Restarting Idle Connections". + */ + if (V_tcp_do_rfc3390) + rw = min(4 * CCV(ccv, t_maxseg), + max(2 * CCV(ccv, t_maxseg), 4380)); + else + rw = CCV(ccv, t_maxseg) * 2; + + CCV(ccv, snd_cwnd) = min(rw, CCV(ccv, snd_cwnd)); +} + /* - * manage congestion signals + * Perform any necessary tasks before we enter congestion recovery. */ -void +static void newreno_cong_signal(struct cc_var *ccv, uint32_t type) { u_int win; @@ -183,11 +204,9 @@ newreno_cong_signal(struct cc_var *ccv, uint32_t type) } /* - * decrease the cwnd in response to packet loss or a transmit timeout. - * th can be null, in which case cwnd will be set according to reno instead - * of new reno. + * Perform any necessary tasks before we exit congestion recovery. */ -void +static void newreno_post_recovery(struct cc_var *ccv) { if (IN_FASTRECOVERY(CCV(ccv, t_flags))) { @@ -209,23 +228,5 @@ newreno_post_recovery(struct cc_var *ccv) } } -/* - * if a connection has been idle for a while and more data is ready to be sent, - * reset cwnd - */ -void -newreno_after_idle(struct cc_var *ccv) -{ - /* - * We have been idle for "a while" and no acks are expected to clock out - * any data we send -- slow start to get ack "clock" running again. - */ - if (V_tcp_do_rfc3390) - CCV(ccv, snd_cwnd) = min(4 * CCV(ccv, t_maxseg), - max(2 * CCV(ccv, t_maxseg), 4380)); - else - CCV(ccv, snd_cwnd) = CCV(ccv, t_maxseg) * 2; -} - DECLARE_CC_MODULE(newreno, &newreno_cc_algo); diff --git a/sys/netinet/if_ether.c b/sys/netinet/if_ether.c index ee5d8e0..df05c04 100644 --- a/sys/netinet/if_ether.c +++ b/sys/netinet/if_ether.c @@ -323,7 +323,7 @@ retry: } if ((la->la_flags & LLE_VALID) && - ((la->la_flags & LLE_STATIC) || la->la_expire > time_second)) { + ((la->la_flags & LLE_STATIC) || la->la_expire > time_uptime)) { bcopy(&la->ll_addr, desten, ifp->if_addrlen); /* * If entry has an expiry time and it is approaching, @@ -331,7 +331,7 @@ retry: * arpt_down interval. */ if (!(la->la_flags & LLE_STATIC) && - time_second + la->la_preempt > la->la_expire) { + time_uptime + la->la_preempt > la->la_expire) { arprequest(ifp, NULL, &SIN(dst)->sin_addr, IF_LLADDR(ifp)); @@ -351,7 +351,7 @@ retry: goto done; } - renew = (la->la_asked == 0 || la->la_expire != time_second); + renew = (la->la_asked == 0 || la->la_expire != time_uptime); if ((renew || m != NULL) && (flags & LLE_EXCLUSIVE) == 0) { flags |= LLE_EXCLUSIVE; LLE_RUNLOCK(la); @@ -403,7 +403,7 @@ retry: int canceled; LLE_ADDREF(la); - la->la_expire = time_second; + la->la_expire = time_uptime; canceled = callout_reset(&la->la_timer, hz * V_arpt_down, arptimer, la); if (canceled) @@ -713,7 +713,7 @@ match: int canceled; LLE_ADDREF(la); - la->la_expire = time_second + V_arpt_keep; + la->la_expire = time_uptime + V_arpt_keep; canceled = callout_reset(&la->la_timer, hz * V_arpt_keep, arptimer, la); if (canceled) diff --git a/sys/netinet/in.c b/sys/netinet/in.c index 2ec54e2..d8c71f1 100644 --- a/sys/netinet/in.c +++ b/sys/netinet/in.c @@ -1333,7 +1333,7 @@ in_lltable_new(const struct sockaddr *l3addr, u_int flags) * For IPv4 this will trigger "arpresolve" to generate * an ARP request. */ - lle->base.la_expire = time_second; /* mark expired */ + lle->base.la_expire = time_uptime; /* mark expired */ lle->l3_addr4 = *(const struct sockaddr_in *)l3addr; lle->base.lle_refcnt = 1; LLE_LOCK_INIT(&lle->base); diff --git a/sys/netinet/ip6.h b/sys/netinet/ip6.h index 3fb08a7..c833a3d 100644 --- a/sys/netinet/ip6.h +++ b/sys/netinet/ip6.h @@ -263,7 +263,7 @@ struct ip6_frag { /* * IP6_EXTHDR_CHECK ensures that region between the IP6 header and the * target header (including IPv6 itself, extension headers and - * TCP/UDP/ICMP6 headers) are continuous. KAME requires drivers + * TCP/UDP/ICMP6 headers) are contiguous. KAME requires drivers * to store incoming data into one internal mbuf or one or more external * mbufs(never into two or more internal mbufs). Thus, the third case is * supposed to never be matched but is prepared just in case. diff --git a/sys/netinet/tcp_input.c b/sys/netinet/tcp_input.c index 8fb9a52..4dbc038 100644 --- a/sys/netinet/tcp_input.c +++ b/sys/netinet/tcp_input.c @@ -378,6 +378,8 @@ cc_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) tp->t_dupacks = 0; tp->t_bytes_acked = 0; EXIT_RECOVERY(tp->t_flags); + tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 / + tp->t_maxseg) * tp->t_maxseg; tp->snd_cwnd = tp->t_maxseg; break; case CC_RTO_ERR: diff --git a/sys/netinet/tcp_output.c b/sys/netinet/tcp_output.c index bf42dac..b7ec6f7 100644 --- a/sys/netinet/tcp_output.c +++ b/sys/netinet/tcp_output.c @@ -148,7 +148,7 @@ tcp_output(struct tcpcb *tp) { struct socket *so = tp->t_inpcb->inp_socket; long len, recwin, sendwin; - int off, flags, error, rw; + int off, flags, error; struct mbuf *m; struct ip *ip = NULL; struct ipovly *ipov = NULL; @@ -182,37 +182,8 @@ tcp_output(struct tcpcb *tp) * to send, then transmit; otherwise, investigate further. */ idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); - if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) { - /* - * If we've been idle for more than one retransmit - * timeout the old congestion window is no longer - * current and we have to reduce it to the restart - * window before we can transmit again. - * - * The restart window is the initial window or the last - * CWND, whichever is smaller. - * - * This is done to prevent us from flooding the path with - * a full CWND at wirespeed, overloading router and switch - * buffers along the way. - * - * See RFC5681 Section 4.1. "Restarting Idle Connections". - */ - if (V_tcp_do_rfc3390) - rw = min(4 * tp->t_maxseg, - max(2 * tp->t_maxseg, 4380)); -#ifdef INET6 - else if ((isipv6 ? in6_localaddr(&tp->t_inpcb->in6p_faddr) : - in_localaddr(tp->t_inpcb->inp_faddr))) -#else - else if (in_localaddr(tp->t_inpcb->inp_faddr)) -#endif - rw = V_ss_fltsz_local * tp->t_maxseg; - else - rw = V_ss_fltsz * tp->t_maxseg; - - tp->snd_cwnd = min(rw, tp->snd_cwnd); - } + if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) + cc_after_idle(tp); tp->t_flags &= ~TF_LASTIDLE; if (idle) { if (tp->t_flags & TF_MORETOCOME) { diff --git a/sys/netinet/tcp_timer.c b/sys/netinet/tcp_timer.c index 2748e64..40ee4d2 100644 --- a/sys/netinet/tcp_timer.c +++ b/sys/netinet/tcp_timer.c @@ -567,7 +567,7 @@ tcp_timer_rexmt(void * xtp) */ tp->t_rtttime = 0; - cc_cong_signal(tp, 0, CC_RTO); + cc_cong_signal(tp, NULL, CC_RTO); (void) tcp_output(tp); |
