Index: share/man/man4/Makefile =================================================================== --- share/man/man4/Makefile +++ share/man/man4/Makefile @@ -87,6 +87,7 @@ cc_cdg.4 \ cc_chd.4 \ cc_cubic.4 \ + cc_dctcp.4 \ cc_hd.4 \ cc_htcp.4 \ cc_newreno.4 \ Index: share/man/man4/cc_dctcp.4 =================================================================== --- /dev/null +++ share/man/man4/cc_dctcp.4 @@ -0,0 +1,269 @@ +.\" +.\" Copyright (c) 2014 Midori Kato +.\" Copyright (c) 2014 The FreeBSD Foundation +.\" All rights reserved. +.\" +.\" Portions of this documentation were written at Keio University, Japan. +.\" +.\" 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. +.\" +.\" $FreeBSD$ +.\" +.Dd Aug 29, 2014 +.Dt CC_DCTCP 4 +.Os +.Sh NAME +.Nm cc_dctcp +.Nd DCTCP Congestion Control Algorithm +.Sh DESCRIPTION +The DCTCP congestion control algorithm is designed to provide short latency +for flows while enabling high throughput for concurrent data transfers in +a data center network. +It leverages ECN (Explicit Congestion Notification) in the network. +DCTCP identifies the depth of congestion from the proportion of feedbacks with +ECN mark. +In terms of congestion window control, DCTCP behaves differently from classic +ECN defined in RFC3168. +.Pp +The congestion window is increased/decreased based on the NewReno congestion +control equations. +DCTCP calculates a fraction of ECE marks in the congestion window +while increasing the congestion window. +When updating the congestion window, DCTCP uses the fraction, whose +window reduction ratio is always equal or less than 1/2. +Only when all packets are marked, congestion window is halved. +.Pp +In order to keep the accuracy of the ECN marked fraction, a DCTCP receiver +mirrors back incoming (or missing) CE marks by setting (or clearing) ECE marks +in transition to or from CE marks. This feedback methodology is also adopted +when the receiver uses delayed ACK. +.Pp +FreeBSD DCTCP implementation includes two minor modifications for the one-sided +deployment. Considering the situation that DCTCP is used as sender and classic +ECN is used as receiver, DCTCP sets the CWR flag as the reaction to the ECE +flag. In addition, when classic ECN is used as sender and DCTCP is used as +receiver, DCTCP avoids to mirror back ACKs only when the CWR flag is +set in the incoming packet. +.Pp +The other specifications are based on the paper and Internet Draft referenced +in the +.Sx SEE ALSO +section below. +.Sh MIB Variables +The algorithm exposes the following tunable variables in the +.Va net.inet.tcp.cc.dctcp +branch of the +.Xr sysctl 3 +MIB: +.Bl -tag -width ".Va alpha" +.It Va alpha +An initial estimator of the congestion on the link. +Default is 0. +.It Va dctcp_shift_g +An estimation gain in the alpha calculation +Default is 16. +.It Va slowstart +A trigger to halve congestion window after slow start. +Default does nothing to halve window. +.Sh SEE ALSO +.Xr cc_chd 4 , +.Xr cc_cubic 4 , +.Xr cc_hd 4 , +.Xr cc_htcp 4 , +.Xr cc_newreno 4 , +.Xr cc_vegas 4 , +.Xr mod_cc 4 , +.Xr tcp 4 , +.Xr mod_cc 9 +.Rs +.%A "Mohammad Alizadeh" +.%A "Albert Greenberg" +.%A "David A. Maltz" +.%A "Jitendra Padhye" +.%A "Parveen Patel" +.%A "Balaji Prabhakar" +.%A "Sudipta Sengupta" +.%A "Murari Sridharan" +.%T "Data Center TCP (DCTCP)" +.%U "http://research.microsoft.com/pubs/121386/dctcp-public.pdf" +.%J "ACM SIGCOMM 2010" +.%D "July 2010" +.%P "63-74" +.Re +.Rs +.%A "Stephen Bensley" +.%A "Lars Eggert" +.%A "Dave Thaler" +.%T "Microsoft's Datacenter TCP (DCTCP): TCP Congestion Control for Datacenters" +.%U "http://tools.ietf.org/html/draft-bensley-tcpm-dctcp-01" +.Re +.Sh HISTORY +The +.Nm +congestion control module first appeared in +.Fx 11.0 . +.Pp +The module was first released in 2014 by Midori Kato studying at Keio +University, Japan. +.Sh AUTHORS +.An -nosplit +The +.Nm +congestion control module and this manual page were written by +.An Midori Kato katoon@sfc.wide.ad.jp +and +.An Lars Eggert lars@netapp.com +.\" +.\" Copyright (c) 2014 Midori Kato +.\" Copyright (c) 2014 The FreeBSD Foundation +.\" All rights reserved. +.\" +.\" Portions of this documentation were written at Keio University, Japan. +.\" +.\" 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. +.\" +.\" $FreeBSD$ +.\" +.Dd June 30, 2013 +.Dt CC_DCTCP 4 +.Os +.Sh NAME +.Nm cc_dctcp +.Nd DCTCP Congestion Control Algorithm +.Sh DESCRIPTION +The DCTCP congestion control algorithm is designed to provide short latency +for flows while enabling high throughput for concurrent data transfers in +a data center network. +It leverages ECN (Explicit Congestion Notification) in the network. +DCTCP identifies the depth of congestion from the proportion of feedbacks with +ECN mark. +In terms of congestion window control, DCTCP behaves differently from classic +ECN defined in RFC3168. +.Pp +The congestion window is increased/decreased based on the NewReno congestion +control equations. +DCTCP calculates a fraction of ECE marks in the congestion window +while increasing the congestion window. +When updating the congestion window, DCTCP uses the fraction, whose +window reduction ratio is always equal or less than 1/2. +Only when all packets are marked, congestion window is halved. +.Pp +In order to keep the accuracy of the ECN marked fraction, a DCTCP receiver +mirrors back incoming (or missing) CE marks by setting (or clearing) ECE marks +in transition to or from CE marks. This feedback methodology is also adopted +when the receiver uses delayed ACK. +.Pp +FreeBSD DCTCP implementation includes two minor modifications for the one-sided +deployment. Considering the situation that DCTCP is used as sender and classic +ECN is used as receiver, DCTCP sets the CWR flag as the reaction to the ECE +flag. In addition, when classic ECN is used as sender and DCTCP is used as +receiver, DCTCP avoids to mirror back ACKs only when the CWR flag is +set in the incoming packet. +.Pp +The other specifications are based on the paper and Internet Draft referenced +in the +.Sx SEE ALSO +section below. +.Sh MIB Variables +The algorithm exposes the following tunable variables in the +.Va net.inet.tcp.cc.dctcp +branch of the +.Xr sysctl 3 +MIB: +.Bl -tag -width ".Va alpha" +.It Va alpha +An initial estimator of the congestion on the link. +Default is 0. +.It Va dctcp_shift_g +An estimation gain in the alpha calculation +Default is 16. +.It Va slowstart +A trigger to halve congestion window after slow start. +Default does nothing to halve window. + +.Sh SEE ALSO +.Xr cc_chd 4 , +.Xr cc_cubic 4 , +.Xr cc_hd 4 , +.Xr cc_htcp 4 , +.Xr cc_newreno 4 , +.Xr cc_vegas 4 , +.Xr mod_cc 4 , +.Xr tcp 4 , +.Xr mod_cc 9 +.Rs +.%A "Mohammad Alizadeh" +.%A "Albert Greenberg" +.%A "David A. Maltz" +.%A "Jitendra Padhye" +.%A "Parveen Patel" +.%A "Balaji Prabhakar" +.%A "Sudipta Sengupta" +.%A "Murari Sridharan" +.%T "Data Center TCP (DCTCP)" +.%U "http://research.microsoft.com/pubs/121386/dctcp-public.pdf" +.%J "ACM SIGCOMM 2010" +.%D "July 2010" +.%P "63-74" +.Re +.Rs +.%A "Stephen Bensley" +.%A "Lars Eggert" +.%A "Dave Thaler" +.%T "Microsoft's Datacenter TCP (DCTCP): TCP Congestion Control for Datacenters" +.%U "http://tools.ietf.org/html/draft-bensley-tcpm-dctcp-01" +.Re +.Sh HISTORY +The +.Nm +congestion control module first appeared in +.Fx 11.0 . +.Pp +The module was first released in 2014 by Midori Kato studying at Keio +University, Japan. +.Sh AUTHORS +.An -nosplit +The +.Nm +congestion control module and this manual page were written by +.An Midori Kato katoon@sfc.wide.ad.jp +and +.An Lars Eggert lars@netapp.com Index: share/man/man4/mod_cc.4 =================================================================== --- share/man/man4/mod_cc.4 +++ share/man/man4/mod_cc.4 @@ -77,6 +77,7 @@ .Sh SEE ALSO .Xr cc_chd 4 , .Xr cc_cubic 4 , +.Xr cc_dctcp 4 , .Xr cc_hd 4 , .Xr cc_htcp 4 , .Xr cc_newreno 4 , Index: sys/modules/cc/Makefile =================================================================== --- sys/modules/cc/Makefile +++ sys/modules/cc/Makefile @@ -3,6 +3,7 @@ SUBDIR= cc_cdg \ cc_chd \ cc_cubic \ + cc_dctcp \ cc_hd \ cc_htcp \ cc_vegas Index: sys/modules/cc/cc_dctcp/Makefile =================================================================== --- /dev/null +++ sys/modules/cc/cc_dctcp/Makefile @@ -0,0 +1,7 @@ +# $FreeBSD$ + +.PATH: ${.CURDIR}/../../../netinet/cc +KMOD= cc_dctcp +SRCS= cc_dctcp.c + +.include @@ -0,0 +1,7 @@ +# $FreeBSD$ + +.PATH: ${.CURDIR}/../../../netinet/cc +KMOD= cc_dctcp +SRCS= cc_dctcp.c + +.include Index: sys/netinet/cc.h =================================================================== --- sys/netinet/cc.h +++ sys/netinet/cc.h @@ -143,6 +143,13 @@ /* Called when data transfer resumes after an idle period. */ void (*after_idle)(struct cc_var *ccv); + /* Called for an additional ECN processing apart from RFC3168. */ + int (*ecnpkt_handler)(struct cc_var *ccv, uint8_t iptos, int cwr, + int is_delayack); + + /* Called when the host marks ECN capable transmission (ECT). */ + int (*ect_handler)(struct cc_var *ccv); + STAILQ_ENTRY (cc_algo) entries; }; Index: sys/netinet/cc/cc_dctcp.c =================================================================== --- /dev/null +++ sys/netinet/cc/cc_dctcp.c @@ -0,0 +1,482 @@ +/*- + * Copyright (c) 2007-2008 + * Swinburne University of Technology, Melbourne, Australia + * Copyright (c) 2009-2010 Lawrence Stewart + * Copyright (c) 2014 Midori Kato + * Copyright (c) 2014 The FreeBSD Foundation + * All rights reserved. + * + * 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 DCTCP algorithm for FreeBSD, based on + * "Data Center TCP (DCTCP)" by M. Alizadeh, A. Greenberg, D. A. Maltz, + * J. Padhye, P. Patel, B. Prabhakar, S. Sengupta, and M. Sridharan., + * in ACM Conference on SIGCOMM 2010, New York, USA, + * Originally released as the contribution of Microsoft Research project. + */ + +#include +__FBSDID("$FreeBSD$"); + +#include +#include +#include +#include +#include +#include +#include +#include + +#include + +#include +#include +#include +#include +#include + +#include + +#define CAST_PTR_INT(X) (*((int*)(X))) + +static VNET_DEFINE(uint32_t, dctcp_alpha) = 0; +static VNET_DEFINE(uint32_t, dctcp_shift_g) = 4; +static VNET_DEFINE(uint32_t, dctcp_slowstart) = 0; +#define V_dctcp_alpha VNET(dctcp_alpha) +#define V_dctcp_shift_g VNET(dctcp_shift_g) +#define V_dctcp_slowstart VNET(dctcp_slowstart) + +struct dctcp { + int bytes_ecn; /* # of marked bytes during a RTT */ + int bytes_total; /* # of acked bytes during a RTT */ + int alpha; /* the fraction of marked bytes */ + int ce_prev; /* CE state of the last segment */ + int save_sndnxt; /* end sequence number of the current window */ + int ece_curr; /* ECE flag in this segment */ + int ece_prev; /* ECE flag in the last segment */ + uint32_t num_cong_events; /* # of congestion events */ +}; + +static MALLOC_DEFINE(M_dctcp, "dctcp data", + "Per connection data required for the dctcp algorithm"); + +static void dctcp_ack_received(struct cc_var *ccv, uint16_t type); +static void dctcp_after_idle(struct cc_var *ccv); +static void dctcp_cb_destroy(struct cc_var *ccv); +static int dctcp_cb_init(struct cc_var *ccv); +static void dctcp_cong_signal(struct cc_var *ccv, uint32_t type); +static void dctcp_conn_init(struct cc_var *ccv); +static void dctcp_post_recovery(struct cc_var *ccv); +static int dctcp_ecnpkt_handler(struct cc_var *ccv, uint8_t iptos, int cwr, + int is_delayack); +static int dctcp_ect_handler(struct cc_var *ccv); +static void dctcp_update_alpha(struct cc_var *ccv); + +struct cc_algo dctcp_cc_algo = { + .name = "dctcp", + .ack_received = dctcp_ack_received, + .cb_destroy = dctcp_cb_destroy, + .cb_init = dctcp_cb_init, + .cong_signal = dctcp_cong_signal, + .conn_init = dctcp_conn_init, + .post_recovery = dctcp_post_recovery, + .ecnpkt_handler = dctcp_ecnpkt_handler, + .after_idle = dctcp_after_idle, + .ect_handler = dctcp_ect_handler, +}; + +static void +dctcp_ack_received(struct cc_var *ccv, uint16_t type) +{ + struct dctcp *dctcp_data; + int bytes_acked = 0; + + dctcp_data = ccv->cc_data; + + /* + * DCTCP doesn't regard with ECN as a congestion. + * Thus, DCTCP always executes the ACK processing out + * of congestion recovery. + */ + if (IN_CONGRECOVERY(CCV(ccv, t_flags))) { + EXIT_CONGRECOVERY(CCV(ccv, t_flags)); + newreno_cc_algo.ack_received(ccv, type); + ENTER_CONGRECOVERY(CCV(ccv, t_flags)); + } else + newreno_cc_algo.ack_received(ccv, type); + + /* Updates the fraction of marked bytes. */ + if (CCV(ccv, t_flags) & TF_ECN_PERMIT) { + + if (type == CC_DUPACK) + bytes_acked = CCV(ccv, t_maxseg); + + if (type == CC_ACK) + bytes_acked = ccv->bytes_this_ack; + + /* Update total bytes. */ + dctcp_data->bytes_total += bytes_acked; + + /* Update total marked bytes. */ + if (dctcp_data->ece_curr) { + if (!dctcp_data->ece_prev + && bytes_acked > CCV(ccv, t_maxseg)) { + dctcp_data->bytes_ecn += + (bytes_acked - CCV(ccv, t_maxseg)); + } else + dctcp_data->bytes_ecn += bytes_acked; + dctcp_data->ece_prev = 1; + } else { + if (dctcp_data->ece_prev + && bytes_acked > CCV(ccv, t_maxseg)) + dctcp_data->bytes_ecn += CCV(ccv, t_maxseg); + dctcp_data->ece_prev = 0; + } + dctcp_data->ece_curr = 0; + + /* + * Update the fraction of marked bytes at the end of + * current window size. + */ + if ((IN_FASTRECOVERY(CCV(ccv, t_flags)) && + SEQ_GEQ(ccv->curack, CCV(ccv, snd_recover))) || + (!IN_FASTRECOVERY(CCV(ccv, t_flags)) && + SEQ_GT(ccv->curack, dctcp_data->save_sndnxt))) + dctcp_update_alpha(ccv); + } +} + +static void +dctcp_after_idle(struct cc_var *ccv) +{ + struct dctcp *dctcp_data; + + dctcp_data = ccv->cc_data; + + /* Initialize internal parameters after idle time */ + dctcp_data->bytes_ecn = 0; + dctcp_data->bytes_total = 0; + dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); + dctcp_data->alpha = V_dctcp_alpha; + dctcp_data->ece_curr = 0; + dctcp_data->ece_prev = 0; + dctcp_data->num_cong_events = 0; + + dctcp_cc_algo.after_idle = newreno_cc_algo.after_idle; +} + +static void +dctcp_cb_destroy(struct cc_var *ccv) +{ + + if (ccv->cc_data != NULL) + free(ccv->cc_data, M_dctcp); +} + +static int +dctcp_cb_init(struct cc_var *ccv) +{ + struct dctcp *dctcp_data; + + dctcp_data = malloc(sizeof(struct dctcp), M_dctcp, M_NOWAIT|M_ZERO); + if (dctcp_data == NULL) + return (ENOMEM); + + /* Initialize some key variables with sensible defaults. */ + dctcp_data->bytes_ecn = 0; + dctcp_data->bytes_total = 0; + /* + * When alpha is set to 0 in the beggining, DCTCP sender transfers as + * much data as possible until the value converges which may expand the + * queueing delay at the switch. When alpha is set to 1, queueing delay + * is kept small. + * Throughput-sensitive applications should have alpha = 0 + * Latency-sensitive applications should have alpha = 1 + * + * Note: DCTCP draft suggests initial alpha to be 1 but we've decided to + * keep it 0. + */ + dctcp_data->alpha = V_dctcp_alpha; + dctcp_data->save_sndnxt = 0; + dctcp_data->ce_prev = 0; + dctcp_data->ece_curr = 0; + dctcp_data->ece_prev = 0; + dctcp_data->num_cong_events = 0; + + ccv->cc_data = dctcp_data; + return (0); +} + +/* + * Perform any necessary tasks before we enter congestion recovery. + */ +static void +dctcp_cong_signal(struct cc_var *ccv, uint32_t type) +{ + struct dctcp *dctcp_data; + u_int win, mss; + + dctcp_data = ccv->cc_data; + win = 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))) { + CCV(ccv, snd_ssthresh) = mss * + max(win / 2 / mss, 2); + dctcp_data->num_cong_events++; + } else { + /* cwnd has already updated as congestion + * recovery. Reverse cwnd value using + * snd_cwnd_prev and recalculate snd_ssthresh + */ + win = CCV(ccv, snd_cwnd_prev); + CCV(ccv, snd_ssthresh) = + max(win / 2 / mss, 2) * mss; + } + ENTER_RECOVERY(CCV(ccv, t_flags)); + } + break; + case CC_ECN: + /* + * Save current snd_cwnd when the host encounters both + * congestion recovery and fast recovery. + */ + CCV(ccv, snd_cwnd_prev) = win; + if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { + if (V_dctcp_slowstart && + dctcp_data->num_cong_events++ == 0) { + CCV(ccv, snd_ssthresh) = + mss * max(win / 2 / mss, 2); + dctcp_data->alpha = 1024; + dctcp_data->bytes_ecn = 0; + dctcp_data->bytes_total = 0; + dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); + } else + CCV(ccv, snd_ssthresh) = max((win - ((win * + dctcp_data->alpha) >> 11)) / mss, 2) * mss; + CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); + ENTER_CONGRECOVERY(CCV(ccv, t_flags)); + } + dctcp_data->ece_curr = 1; + break; + case CC_RTO: + if (CCV(ccv, t_flags) & TF_ECN_PERMIT) { + CCV(ccv, t_flags) |= TF_ECN_SND_CWR; + dctcp_update_alpha(ccv); + dctcp_data->save_sndnxt += CCV(ccv, t_maxseg); + dctcp_data->num_cong_events++; + } + break; + } +} + +static void +dctcp_conn_init(struct cc_var *ccv) +{ + struct dctcp *dctcp_data; + + dctcp_data = ccv->cc_data; + + if (CCV(ccv, t_flags) & TF_ECN_PERMIT) + dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); +} + +/* + * Perform any necessary tasks before we exit congestion recovery. + */ +static void +dctcp_post_recovery(struct cc_var *ccv) +{ + + dctcp_cc_algo.post_recovery = newreno_cc_algo.post_recovery; + + if (CCV(ccv, t_flags) & TF_ECN_PERMIT) + dctcp_update_alpha(ccv); +} + +/* + * Execute an additional ECN processing using ECN field in IP header and the CWR + * bit in TCP header. + * + * Return: + * 1 - Delayed ACK disabled + * 0 - Delayed ACK not disabled + */ +static int +dctcp_ecnpkt_handler(struct cc_var *ccv, uint8_t iptos, int cwr, int is_delayack) +{ + struct dctcp *dctcp_data; + int ret; + + dctcp_data = ccv->cc_data; + ret = 0; + + switch (iptos & IPTOS_ECN_MASK) { + case IPTOS_ECN_CE: + if (!dctcp_data->ce_prev && is_delayack) + ret = 1; + dctcp_data->ce_prev = 1; + CCV(ccv, t_flags) |= TF_ECN_SND_ECE; + break; + case IPTOS_ECN_ECT0: + case IPTOS_ECN_ECT1: + if (dctcp_data->ce_prev && is_delayack) + ret = 1; + CCV(ccv, t_flags) &= ~TF_ECN_SND_ECE; + dctcp_data->ce_prev = 0; + break; + } + + /* + * DCTCP receiver send an ACK immediately + * - when the CE state in between this segment + * and the last segment is not the same + * - when this segment sets the CWR flag + */ + if (cwr && is_delayack) + ret = 0; + + return (ret); +} + +static int +dctcp_ect_handler(struct cc_var *ccv) +{ + /* DCTCP always marks ECT */ + return (1); +} + +/* + * Update the fraction of marked bytes represented as 'alpha'. + * Also initialize several internal parameters at the end of this function. + */ +static void +dctcp_update_alpha(struct cc_var *ccv) +{ + struct dctcp *dctcp_data; + int alpha_prev; + + dctcp_data = ccv->cc_data; + alpha_prev = dctcp_data->alpha; + dctcp_data->bytes_total = max(dctcp_data->bytes_total, 1); + + /* + * Update alpha: alpha = (1 - g) * alpha + g * F. + * Here: + * g is weight factor + * recommaded to be set to 1/16 + * small g = slow convergence between competitive DCTCP flows + * large g = impacts low utilization of bandwidth at switches + * F is fraction of marked segments in last RTT + * updated every RTT + * Alpha must be round to 0 - 1024. + * XXXMIDORI Is more fine-grained alpha necessary? + */ + dctcp_data->alpha = min(alpha_prev - (alpha_prev >> V_dctcp_shift_g) + + (dctcp_data->bytes_ecn << (10 - V_dctcp_shift_g)) / + dctcp_data->bytes_total, 1024); + + /* Initialize internal parameters for next alpha calculation */ + dctcp_data->bytes_ecn = 0; + dctcp_data->bytes_total = 0; + dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); +} + +static int +dctcp_alpha_handler(SYSCTL_HANDLER_ARGS) +{ + uint32_t new; + int error; + + new = V_dctcp_alpha; + error = sysctl_handle_int(oidp, &new, 0, req); + if (error == 0 && req->newptr != NULL) { + if (CAST_PTR_INT(req->newptr) > 1) + error = EINVAL; + else + V_dctcp_alpha = new; + } + + return (error); +} + +static int +dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS) +{ + uint32_t new; + int error; + + new = V_dctcp_shift_g; + error = sysctl_handle_int(oidp, &new, 0, req); + if (error == 0 && req->newptr != NULL) { + if (CAST_PTR_INT(req->newptr) > 1) + error = EINVAL; + else + V_dctcp_shift_g = new; + } + + return (error); +} + +static int +dctcp_slowstart_handler(SYSCTL_HANDLER_ARGS) +{ + uint32_t new; + int error; + + new = V_dctcp_slowstart; + error = sysctl_handle_int(oidp, &new, 0, req); + if (error == 0 && req->newptr != NULL) { + if (CAST_PTR_INT(req->newptr) > 1) + error = EINVAL; + else + V_dctcp_slowstart = new; + } + + return (error); +} + +SYSCTL_DECL(_net_inet_tcp_cc_dctcp); +SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, dctcp, CTLFLAG_RW, NULL, + "dctcp congestion control related settings"); + +SYSCTL_VNET_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, alpha, + CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_alpha), 0, + &dctcp_alpha_handler, + "IU", "dctcp alpha parameter"); + +SYSCTL_VNET_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, shift_g, + CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_shift_g), 4, + &dctcp_shift_g_handler, + "IU", "dctcp shift parameter"); + +SYSCTL_VNET_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, slowstart, + CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_slowstart), 0, + &dctcp_slowstart_handler, + "IU", "half CWND reduction after the first slow start"); + +DECLARE_CC_MODULE(dctcp, &dctcp_cc_algo); Index: sys/netinet/tcp_input.c =================================================================== --- sys/netinet/tcp_input.c +++ sys/netinet/tcp_input.c @@ -492,6 +492,19 @@ (tlen <= tp->t_maxopd) && \ (V_tcp_delack_enabled || (tp->t_flags & TF_NEEDSYN))) +static void inline +cc_ecnpkt_handler(struct tcpcb *tp, struct tcphdr *th, uint8_t iptos, int tlen) +{ + INP_WLOCK_ASSERT(tp->t_inpcb); + + if (CC_ALGO(tp)->ecnpkt_handler != NULL) { + if (CC_ALGO(tp)->ecnpkt_handler(tp->ccv, iptos, + (th->th_flags & TH_CWR), DELAY_ACK(tp, tlen))) { + tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); + } + } +} + /* * TCP input handling is split into multiple parts: * tcp6_input is a thin wrapper around tcp_input for the extended @@ -1523,6 +1536,10 @@ TCPSTAT_INC(tcps_ecn_ect1); break; } + + /* Process a packet differently from RFC3168. */ + cc_ecnpkt_handler(tp, th, iptos, tlen); + /* Congestion experienced. */ if (thflags & TH_ECE) { cc_cong_signal(tp, th, CC_ECN); Index: sys/netinet/tcp_output.c =================================================================== --- sys/netinet/tcp_output.c +++ sys/netinet/tcp_output.c @@ -161,6 +161,24 @@ CC_ALGO(tp)->after_idle(tp->ccv); } +static int inline +cc_ect_handler(struct tcpcb *tp) +{ + INP_WLOCK_ASSERT(tp->t_inpcb); + + if (CC_ALGO(tp)->ect_handler != NULL) { + if (CC_ALGO(tp)->ect_handler(tp->ccv)) + return (1); + else + return (0); + } + /* + * Currently, only dctcp needs/implements ect_handler. So if a module + * does not implement it, return SUCCESS. + */ + return (1); +} + /* * Tcp output routine: figure out what should be sent and send it. */ @@ -965,9 +983,15 @@ * If the peer has ECN, mark data packets with * ECN capable transmission (ECT). * Ignore pure ack packets, retransmissions and window probes. + * Mark data packet with ECN capable transmission (ECT) + * when CC_ALGO meets specific condition. + * Or, if the peer has ECN, mark data packets with ECT + * (RFC 3168). Ignore pure ack packets, retransmissions + * and window probes. */ - if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && - !((tp->t_flags & TF_FORCEDATA) && len == 1)) { + if (cc_ect_handler(tp) && + (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) + && !((tp->t_flags & TF_FORCEDATA) && len == 1))) { #ifdef INET6 if (isipv6) ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);