Re: [PATCH v5 28/28] x86/fpu/amx: Clear the AMX state when appropriate

From: Peter Zijlstra
Date: Thu May 27 2021 - 07:56:51 EST


On Mon, May 24, 2021 at 02:24:54PM -0400, Len Brown wrote:
> On Mon, May 24, 2021 at 1:39 PM Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
>
> > >> might be considered nasty.
>
> > > I'm not excited about burdening the generic idle path with a CPU
> > > feature specific check that would need to be checked on every idle
> > > entry.
> >
> > Me neither.
> >
> > But, the check itself should be cheap. A cpu_feature_enabled(AMX) check
> > will eliminate even the cost of a branch on systems without AMX. You
> > could probably even get fancy and also use a static branch that doesn't
> > get enabled until the first AMX user shows up.
>
> It isn't just the hardware run-time cost.
> It is the source code complexity.
> That code is 100% generic.

Why can't this live in intel_idle.c ? We had to pull out
CPUIDLE_FLAG_TLB_FLUSH because leave_mm() requires RCU, but afaict both
TILERELEASE and XRESTOR do not and could be added to intel_idle_tile(),
which can be used in XXX_cstates[] for the relevant chips instead of
intel_idle() for C6 and up.

intel_idle_tile(args)
{
bool has_tile = // something XINUSE

if (has_tile)
TILERELEASE

intel_idle(args...);

if (has_tile)
// something XRESTOR
}

Hmm?