Re: [PATCH 1/8] perf/x86/intel: Fix memory corruption

[Peter Zijlstra]

On Fri, Mar 22, 2019 at 12:04:55PM -0700, Stephane Eranian wrote:
> On Thu, Mar 21, 2019 at 10:51 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >
> > On Thu, 21 Mar 2019, Stephane Eranian wrote:
> > > On Thu, Mar 21, 2019 at 9:45 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> > > >
> > > > On Thu, 21 Mar 2019, Peter Zijlstra wrote:
> > > > > Subject: perf/x86/intel: Initialize TFA MSR
> > > > >
> > > > > Stephane reported that we don't initialize the TFA MSR, which could lead
> > > > > to trouble if the RESET value is not 0 or on kexec.
> > > >
> > > > That sentence doesn't parse.
> > > >
> > > >   Stephane reported that the TFA MSR is not initialized by the kernel, but
> > > >   the TFA bit could set by firmware or as a leftover from a kexec, which
> > > >   makes the state inconsistent.
> > > >
> > > Correct. This is what I meant.
> > > The issue is what does the kernel guarantee when it boots?
> > >
> > > I see:
> > > static bool allow_tsx_force_abort = true;
> > >
> > > Therefore you must ensure the MSR is set to reflect that state on boot.
> > > So you have to force it to that value to be in sync which is what your
> > > new patch is doing.
> >
> > The initial state should be that the MSR TFA bit is 0. The software state
> > is a different beast.
> >
> > allow_tsx_force_abort
> >
> >   false                 Do not set MSR TFA bit (Make TSX work with PMC3) and
> >                         exclude PMC3 from being used.
> >
> >   true                  Set the MSR TFA bit when PMC3 is used by perf, clear it
> >                         when PMC3 is not longer in use.
> >
> I would expect this description to be included in the source code where the
> allow_tsx_force_abort variable is defined

That part is easy I suppose.

> and somewhere in the kernel Documentation because it is not trivial to
> understand what the control actually does and the guarantees you have
> when you toggle it.

But here we seem to be sorely lacking, that is, there is no sysfs/perf
documentation at all.

In any case, feel free to send a patch :-)