Date: Tue, 20 Jan 2015 11:07:53 -0500 From: John Baldwin <jhb@freebsd.org> To: Hans Petter Selasky <hps@selasky.org> Cc: Konstantin Belousov <kostikbel@gmail.com>, "svn-src-head@freebsd.org" <svn-src-head@freebsd.org>, Adrian Chadd <adrian@freebsd.org>, "src-committers@freebsd.org" <src-committers@freebsd.org>, "svn-src-all@freebsd.org" <svn-src-all@freebsd.org> Subject: Re: svn commit: r277213 - in head: share/man/man9 sys/kern sys/ofed/include/linux sys/sys Message-ID: <201501201107.53553.jhb@freebsd.org> In-Reply-To: <54BE21F0.6010602@selasky.org> References: <201501151532.t0FFWV2Y037455@svn.freebsd.org> <20150120090057.GD42409@kib.kiev.ua> <54BE21F0.6010602@selasky.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tuesday, January 20, 2015 4:37:52 am Hans Petter Selasky wrote:
> On 01/20/15 10:00, Konstantin Belousov wrote:
> > On Tue, Jan 20, 2015 at 08:58:34AM +0100, Hans Petter Selasky wrote:
> >> On 01/20/15 08:51, Konstantin Belousov wrote:
> >>> On Tue, Jan 20, 2015 at 05:30:25AM +0100, Hans Petter Selasky wrote:
> >>>> On 01/19/15 22:59, Adrian Chadd wrote:
> >>>>> Hi,
> >>>>>
> >>>>> Would you please check what the results of this are with CPU specific
> >>>>> callwheels?
> >>>>>
> >>>>> I'm doing some 10+ gig traffic testing on -HEAD with RSS enabled (on
> >>>>> ixgbe) and with this setup, the per-CPU TCP callwheel stuff is
> >>>>> enabled. But all the callwheels are now back on clock(0) and so is the
> >>>>> lock contention. :(
> >>>>>
> >>>>> Thanks,
> >>>>>
> >>>>
> >>>> Hi,
> >>>>
> >>>> Like stated in the manual page, callout_reset_curcpu/on() does not work
> >>>> with MPSAFE callouts any more!
> >>> I.e. you 'fixed' some undeterminate bugs in callout migration by not
> >>> doing migration at all anymore.
> >>>
> >>>>
> >>>> You need to use callout_init_{mtx,rm,rw} and remove the custom locking
> >>>> inside the callback in the TCP stack to get it working like before!
> >>>
> >>> No, you need to do this, if you think that whole callout KPI must be
> >>> rototiled. It is up to the person who modifies the KPI, to ensure that
> >>> existing code is not broken.
>
> Hi,
>
> It is not very hard to update existing callout clients and you can do it
> too, if you need the extra bits of performance.
>
> Are there more API's than the TCP stack which you think needs an update
> and are performance critical?
>
> >>>
> >>> As I understand, currently we are back to the one-cpu callouts.
> >>> Do other people consider this situation acceptable ?
>
> For the TCP stack - yes, but not for other clients like cv_timedwait()
> and such.
>
> If you think you have a better way to solve the callout problems, please
> tell me! In order for a callout to change its CPU you need a lock to
> protect which CPU the callout is on. Instead of introducing a third lock
> in the callout path, which will be a congestion point, to protect
> against changing the CPU number, I decided that we will use the client's
> mutex and the MPSAFE implies the client doesn't have any mutex. So it
> won't work with callout clients which use the CALLOUT_MPSAFE flag.
> Honestly CALLOUT_MPSAFE should not be used, because it leads to extra
> complexity in the clients catching the race when tearing down the
> callouts and any pending callbacks.
>
> >>
> >> Please read the callout 9 manual page first.
> >
> > Assume I read it. How this changes any of my points above ?
> > """
> > A change in the CPU selection cannot happen if this function is
> > re-scheduled inside a callout function. Else the callback function given
> > by the func argument will be executed on the same CPU like previously
> > done.
> > """
> > You cannot do this without fixing consumers.
> >
>
> The code simply needs an update. It is not broken in any ways - right?
> If it is not broken, fixing it is not that urgent.
This is not at all acceptable. TCP callouts were the largest potential user
of multi-cpu callouts and you've just broken them. Your proposed change to
handle inp locks is not necessarily correct either since dropping the inp lock
inside a callout introduces new races (now callout_stop doesn't have quite the
same semantics as it does for other callout_init_*()). Given this, it seems
that your fix just mostly disabled multi-CPU callouts, so it is not at all
clear that you've actually fixed anything. :(
--
John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201501201107.53553.jhb>