Re: [RFC PATCH 3/3] nSVM: use svm->nested.save to load vmcb12 registers and avoid TOC/TOU races

[Maxim Levitsky]

On Tue, 2021-09-14 at 11:24 +0200, Emanuele Giuseppe Esposito wrote:
> 
> On 14/09/2021 11:12, Maxim Levitsky wrote:
> > On Tue, 2021-09-14 at 12:02 +0300, Maxim Levitsky wrote:
> > > On Tue, 2021-09-14 at 10:20 +0200, Emanuele Giuseppe Esposito wrote:
> > > > On 12/09/2021 12:42, Maxim Levitsky wrote:
> > > > > >    
> > > > > > -	if (!nested_vmcb_valid_sregs(vcpu, &vmcb12->save) ||
> > > > > > +	if (!nested_vmcb_valid_sregs(vcpu, &svm->nested.save) ||
> > > > > >    	    !nested_vmcb_check_controls(vcpu, &svm->nested.ctl)) {
> > > > > If you use a different struct for the copied fields, then it makes
> > > > > sense IMHO to drop the 'control' parameter from nested_vmcb_check_controls,
> > > > > and just use the svm->nested.save there directly.
> > > > > 
> > > > 
> > > > Ok, what you say in patch 2 makes sense to me. I can create a new struct
> > > > vmcb_save_area_cached, but I need to keep nested.ctl because 1) it is
> > > > used also elsewhere, and different fields from the one checked here are
> > > > read/set and 2) using another structure (or the same
> > > 
> > > Yes, keep nested.ctl, since vast majority of the fields are copied I think.
> > 
> > But actually that you mention it, I'll say why not to create vmcb_control_area_cached
> > as well indeed and change the type of svm->nested.save to it. (in a separate patch)
> > 
> > I see what you mean that we modify it a bit (but we shoudn't to be honest) and such, but
> > all of this can be fixed.
> 
> So basically you are proposing:
> 
> struct svm_nested_state {
> 	...
> 	struct vmcb_control_area ctl; // we need this because it is used 
> everywhere, I think
> 	struct vmcb_control_area_cached ctl_cached;
> 	struct vmcb_save_area_cached save_cached;
> 	...
> }
> 
> and then
> 
> if (!nested_vmcb_valid_sregs(vcpu, &svm->nested.save_cached) ||
>      !nested_vmcb_check_controls(vcpu, &svm->nested.ctl_cached)) {
> 
> like that?
> 
> Or do you want to delete nested.ctl completely and just keep the fields 
> actually used in ctl_cached?


I would do it this way:

struct svm_nested_state {
        ...
	/* cached fields from the vmcb12 */
	struct  vmcb_control_area_cached ctl;
	struct  vmcb_save_area_cached save;
        ...
};


Best regards,
     Maxim Levitsky

> 
> 
> Also, note that as I am trying to use vmcb_save_area_cached, it is worth 
> noticing that nested_vmcb_valid_sregs() is also used in 
> svm_set_nested_state(), so it requires some additional little changes.
> 
> Thank you,
> Emanuele
> 
> > The advantage of having vmcb_control_area_cached is that it becomes impossible to use
> > by mistake a non copied field from the guest.
> > 
> > It would also emphasize that this stuff came from the guest and should be treated as
> > a toxic waste.
> > 
> > Note again that this should be done if we agree as a separate patch.
> > 
> > > Best regards,
> > > 	Maxim Levitsky
> > > 
> > > 
> > > > vmcb_save_area_cached) in its place would just duplicate the same fields
> > > > of nested.ctl, creating even more confusion and possible inconsistency.
> > > > 
> > > > Let me know if you disagree.
> > > > 
> > > > Thank you,
> > > > Emanuele
> > > >