Re: [PATCH 4/8] x86/traps: Demand-populate PASID MSR via #GP

[Peter Zijlstra]

On Wed, Sep 22, 2021 at 09:26:10PM +0000, Luck, Tony wrote:
> >> > +static bool fixup_pasid_exception(void)
> >> > +{
> >> > +	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
> >> > +		return false;
> >> > +
> >> > +	return __fixup_pasid_exception();
> >> > +}
> >
> > That is, shouldn't the above at the very least decode the instruction
> > causing the #GP and check it's this ENQCMD thing?
> 
> It can't reliably do that because some other thread in the process may
> have re-written the memory that regs->ip points at (bizarre case, but
> I think Dave Hansen brought it up).

I don't buy that argument, any cross modifying code gets to keep the
pieces in that case.

> So it would just add extra code, and still only be a hint.
> 
> Without the check this sequence is possible:
> 
> 1) Process binds an accelerator (so mm->pasid is set)
> 2) Task in the process executes something other than ENQCMD that gets a #GP
> 3) Kernel says "Oh, mm->pasid is set, I'll initialize the IA32_PASID MSR to see if that fixes it"
> 4) Nope. Re-executing the instruction at step #2 just gives another #GP
> 5) Kernel says "I already set IA32_PASID, so this must be something else ... do regular #GP actions"
> 
> Now if the task catches the signal that results from step #5 and avoids termination, it will have
> IA32_PASID set ... but to the right value should it go on to actually execute ENQCMD at some
> future point.
> 
> So the corner case from not knowing whether this #GP was from ENQCMD or not is harmless.

And all that *really* should be a in a comment near there, because I'm
100% sure I'll get confused and wonder about that very same thing the
next time I see that code.