Date: Sat, 7 Sep 2019 11:33:28 +0000 (UTC) From: Michael Tuexen <tuexen@FreeBSD.org> To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-stable@freebsd.org, svn-src-stable-12@freebsd.org Subject: svn commit: r351994 - in stable/12: share/man/man4 sys/netinet/cc Message-ID: <201909071133.x87BXSRR043314@repo.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: tuexen Date: Sat Sep 7 11:33:27 2019 New Revision: 351994 URL: https://svnweb.freebsd.org/changeset/base/351994 Log: MFC r350403: * Improve input validation of sysctl parameters for DCTPC. * Initialize the alpha parameter to a conservative value (like Linux) * Improve handling of arithmetic. * Improve man-page Obtained from: Richard Scheffenegger Differential Revision: https://reviews.freebsd.org/D20549 Modified: stable/12/share/man/man4/cc_dctcp.4 stable/12/sys/netinet/cc/cc_dctcp.c Directory Properties: stable/12/ (props changed) Modified: stable/12/share/man/man4/cc_dctcp.4 ============================================================================== --- stable/12/share/man/man4/cc_dctcp.4 Sat Sep 7 11:31:05 2019 (r351993) +++ stable/12/share/man/man4/cc_dctcp.4 Sat Sep 7 11:33:27 2019 (r351994) @@ -28,7 +28,7 @@ .\" .\" $FreeBSD$ .\" -.Dd January 12, 2015 +.Dd July 29, 2019 .Dt CC_DCTCP 4 .Os .Sh NAME @@ -60,7 +60,7 @@ In addition, when classic ECN is used as sender and DC 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 +The other specifications are based on the paper and the RFC referenced in the .Sx SEE ALSO section below. @@ -70,16 +70,27 @@ The algorithm exposes the following tunable variables branch of the .Xr sysctl 3 MIB: -.Bl -tag -width ".Va alpha" +.Bl -tag -width ".Va slowstart" .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. +The initial value to estimate the congestion on the link. +The valid range is from 0 to 1024, where 1024 reduces the congestion +window to half, if a CE is observed in the first window and +.Va alpha +could not yet adjust to the congestion level on that path. +Default is 1024. +.It Va shift_g +An estimation gain in the +.Va alpha +calculation. +This influences the responsiveness when adjusting alpha +to the most recent observed window. +Valid range from 0 to 10, the default is 4, resulting in an effective +gain of 1 / ( 2 ^ +.Va shift_g +), or 1/16th. .It Va slowstart -A trigger to halve congestion window after slow start. -Default does nothing to halve window. +A flag if the congestion window should be reduced by one half after slow start. +Valid settings 0 and 1, default 0. .El .Sh SEE ALSO .Xr cc_chd 4 , @@ -108,10 +119,12 @@ Default does nothing to halve window. .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" +.%A "Praveen Balasubramanian" +.%A "Lars Eggert" +.%A "Glenn Judd" +.%T "Data Center TCP (DCTCP): TCP Congestion Control for Data Centers" +.%U "https://tools.ietf.org/html/rfc8257" .Re .Sh HISTORY The Modified: stable/12/sys/netinet/cc/cc_dctcp.c ============================================================================== --- stable/12/sys/netinet/cc/cc_dctcp.c Sat Sep 7 11:31:05 2019 (r351993) +++ stable/12/sys/netinet/cc/cc_dctcp.c Sat Sep 7 11:33:27 2019 (r351994) @@ -56,8 +56,9 @@ __FBSDID("$FreeBSD$"); #include <netinet/cc/cc.h> #include <netinet/cc/cc_module.h> -#define MAX_ALPHA_VALUE 1024 -VNET_DEFINE_STATIC(uint32_t, dctcp_alpha) = 0; +#define DCTCP_SHIFT 10 +#define MAX_ALPHA_VALUE (1<<DCTCP_SHIFT) +VNET_DEFINE_STATIC(uint32_t, dctcp_alpha) = MAX_ALPHA_VALUE; #define V_dctcp_alpha VNET(dctcp_alpha) VNET_DEFINE_STATIC(uint32_t, dctcp_shift_g) = 4; #define V_dctcp_shift_g VNET(dctcp_shift_g) @@ -65,14 +66,14 @@ VNET_DEFINE_STATIC(uint32_t, dctcp_slowstart) = 0; #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 */ + uint32_t bytes_ecn; /* # of marked bytes during a RTT */ + uint32_t bytes_total; /* # of acked bytes during a RTT */ + int alpha; /* the fraction of marked bytes */ + int ce_prev; /* CE state of the last segment */ + tcp_seq 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", @@ -369,18 +370,18 @@ dctcp_update_alpha(struct cc_var *ccv) dctcp_data->bytes_total = max(dctcp_data->bytes_total, 1); /* - * Update alpha: alpha = (1 - g) * alpha + g * F. + * Update alpha: alpha = (1 - g) * alpha + g * M. * 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 + * M is fraction of marked segments in last RTT * updated every RTT * Alpha must be round to 0 - MAX_ALPHA_VALUE. */ - dctcp_data->alpha = min(alpha_prev - (alpha_prev >> V_dctcp_shift_g) + - (dctcp_data->bytes_ecn << (10 - V_dctcp_shift_g)) / + dctcp_data->alpha = ulmin(alpha_prev - (alpha_prev >> V_dctcp_shift_g) + + ((uint64_t)dctcp_data->bytes_ecn << (DCTCP_SHIFT - V_dctcp_shift_g)) / dctcp_data->bytes_total, MAX_ALPHA_VALUE); /* Initialize internal parameters for next alpha calculation */ @@ -398,14 +399,10 @@ dctcp_alpha_handler(SYSCTL_HANDLER_ARGS) new = V_dctcp_alpha; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { - if (new > 1) + if (new > MAX_ALPHA_VALUE) error = EINVAL; - else { - if (new > MAX_ALPHA_VALUE) - V_dctcp_alpha = MAX_ALPHA_VALUE; - else - V_dctcp_alpha = new; - } + else + V_dctcp_alpha = new; } return (error); @@ -420,7 +417,7 @@ dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS) new = V_dctcp_shift_g; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { - if (new > 1) + if (new > DCTCP_SHIFT) error = EINVAL; else V_dctcp_shift_g = new; @@ -454,7 +451,7 @@ SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, dctcp, CTLFLAG SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, alpha, CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_alpha), 0, &dctcp_alpha_handler, - "IU", "dctcp alpha parameter"); + "IU", "dctcp alpha parameter at start of session"); SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, shift_g, CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_shift_g), 4,
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201909071133.x87BXSRR043314>