Re: [PATCH v20 06/14] KVM: arm64: Validate GCS exception lock when emulating ERET

[Leonardo Bras]

On Thu, Sep 03, 2026 at 08:22:44PM +0100, Mark Brown wrote:
> On Thu, Sep 03, 2026 at 04:37:37PM +0100, Leonardo Bras wrote:
> > On Tue, Sep 01, 2026 at 10:47:04PM +0100, Mark Brown wrote:
> 
> > > +	return vcpu_read_sys_reg(vcpu, GCSCR_EL2) & GCSCR_ELx_EXLOCKEN;
> > > +}
> 
> > The above perfectly translates the GCS part of IllegalExceptionReturn().
> 
> > It's a nit, as I suppose there should be no compiler warning on that, but
> > the function should return a bool, and the last return line returns an u64.
> 
> > Maybe adding a "return !!()" would be better?
> 
> There's no need to manually do translations like that in C, the
> conversion of 0 to false and any non-zero value to true when an
> integer is used in a boolean context is in the standard.
> 

Yeah, I am aware the conversion will happen anyway, but I remember someone 
complaining about something like this in the past, that's why I sent as a 
nit. 

> > >  	 * - trying to return to EL1 with HCR_EL2.TGE set
> > > +	 * - GCSCR_ELx.EXLOCKEN is 1 and PSTATE.EXLOCK is 0 when attempting
> > > +	 *   to return from ELx the same EL.
> > >  	 */
> > >  	if (mode == PSR_MODE_EL3t   || mode == PSR_MODE_EL3h ||
> > >  	    mode == 0b00001         || (mode & BIT(1))       ||
> > >  	    (spsr & PSR_MODE32_BIT) ||
> > > +	    kvm_check_illegal_exlock_return(vcpu, spsr) ||
> > >  	    (vcpu_el2_tge_is_set(vcpu) && (mode == PSR_MODE_EL1t ||
> > >  					   mode == PSR_MODE_EL1h))) {
> > >  		u64 mask;
> > 
> > In IllegalExceptionReturn(), the GCS-related clause happens at the end, and 
> > here it happens before the TGE one. Could this cause any weird behavior in 
> > the future?
> > 
> > I get that by doing like this you don't change the last line of the "if", 
> > but I wonder if that could change anything.
> 
> Given that we take the same action regardless of which or clause
> triggers I can't see how it would matter.

I see... well, as long as neither test ever have any collateral effect, I 
think it should not matter, then.

> 
> > > --- a/arch/arm64/kvm/hyp/vhe/switch.c
> > > +++ b/arch/arm64/kvm/hyp/vhe/switch.c
> > > @@ -383,6 +383,10 @@ static bool kvm_hyp_handle_eret(struct kvm_vcpu *vcpu, u64 *exit_code)
> > >  		return false;
> > >  	}
> 
> > > +	/* Push GCS exception lock failures into the slow path */
> > > +	if (kvm_check_illegal_exlock_return(vcpu, spsr))
> > > +		return false;
> 
> > >  	/* If ERETAx fails, take the slow path */
> > >  	if (esr_iss_is_eretax(esr)) {
> > >  		if (!(vcpu_has_ptrauth(vcpu) && kvm_auth_eretax(vcpu, &elr)))
> 
> > IIUC, this function will be called on the __kvm_vcpu_run_vhe() inner loop, 
> > in the cases where the guest exited due to a eret. 
> 
> > What you change here is that in case of an illegal exlock return, it goes 
> > out of the loop and return to host kernel, probably to deal with it in the 
> > mentioned slowpath, the same way the ERETAx entry does.
> 
> > I don't question on this being needed. 
> > I would just like to understand why this is needed here. 
> 
> > This does not seem to be related to nested, as this is called in 
> > __fixup_guest_exit() and not in fixup_nv_guest_exit(). But would not 
> > hardware be responsible for cheking this, then?
> 
> The code is here because it's part of the ERET handling, this should
> only happen for NV as we're not trapping ERET instructions otherwise.

Oh, makes sense!

> We need this because ERETs from vEL2 are handled in software, modulo the
> NV3 fast path mentioned at the top of the function.

Oh, and this is done in __fixup_guest_exit() because vEL2 is not a 
nested guest. It would be it's guests' exit that would be dealt in 
fixup_nv_guest_exit().

Is this correct?

Thanks!
Leo