Re: [RFC 5/9] RISC-V: KVM: Add skeleton support for perf

[Andrew Jones]

On Thu, Nov 24, 2022 at 01:04:15AM -0800, Atish Patra wrote:
> On Wed, Nov 23, 2022 at 5:36 AM Andrew Jones <ajones@xxxxxxxxxxxxxxxx> wrote:
> >
> > ...
> > > > > > -     csr_write(CSR_HCOUNTEREN, -1UL);
> > > > > > +     /* VS should access only TM bit. Everything else should trap */
> > > > > > +     csr_write(CSR_HCOUNTEREN, 0x02);
> > > > >
> > > > > This looks like something that should be broken out into a separate patch
> > > > > with a description of what happens now when guests try to access the newly
> > > > > trapping counter registers. We should probably also create a TM define.
> > > > >
> > > >
> > > > Done.
> > > >
> > >
> > > As we allow cycles & instret for host user space now [1], should we do the same
> > > for guests as well ? I would prefer not to but same user space
> > > software will start to break
> > > they will run inside a guest.
> > >
> > > https://lore.kernel.org/all/20220928131807.30386-1-palmer@xxxxxxxxxxxx/
> > >
> >
> > Yeah, it seems like we should either forbid access to unprivileged users
> > or ensure the numbers include some random noise. For guests, a privileged
> > KVM userspace should need to explicitly request access for them, ensuring
> > that the creation of privileged guests is done by conscious choice.
> >
> 
> If I understand you correctly, you are suggesting we only enable TM
> bit in hcounteren ?

Yeah, and also that I think it'd be nice to revisit this for userspace.

> We also need a mechanism to enable the hcounteren bits from KVM guest userspace.
> 
> I can think of the following approaches.
> 
> 1. The CYCLE, INSTRET enabling can also be via one reg interface.
> 2. We can enable it during first virtual instruction trap if these
> bits in guest scounteren
> are enabled.

Those sound good to me.

Thanks,
drew