Re: [PATCH] KVM: arm64: Fix NULL pointer access issue

[James Clark]

On 01/09/2025 1:24 pm, Marc Zyngier wrote:
> On Mon, 01 Sep 2025 11:36:11 +0100,
> James Clark <james.clark@xxxxxxxxxx> wrote:
> > On 01/09/2025 11:01 am, Yingchao Deng wrote:
> > > When linux is booted in EL1, macro "host_data_ptr()" is a wrapper that
> > > resolves to "&per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)",
> > > is_hyp_mode_available() return false during kvm_arm_init, the per-CPU base
> > > pointer __kvm_nvhe_kvm_arm_hyp_percpu_base[cpu] remains uninitialized.
> > > Consequently, any access via per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)
> > > will result in a NULL pointer.
> > >
> > > Add is_kvm_arm_initialised() condition check to ensure that kvm_arm_init
> > > completes all necessary initialization steps, including init_hyp_mode.
> > >
> > > Fixes: 054b88391bbe2 ("KVM: arm64: Support trace filtering for guests")
> > > Signed-off-by: Yingchao Deng <yingchao.deng@xxxxxxxxxxxxxxxx>
> > > ---
> > > Add a check to prevent accessing uninitialized per-CPU data.
> > > ---
> > >    arch/arm64/kvm/debug.c | 7 ++++---
> > >    1 file changed, 4 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/arch/arm64/kvm/debug.c b/arch/arm64/kvm/debug.c
> > > index 381382c19fe4741980c79b08bbdab6a1bcd825ad..add58056297293b4eb337028773b1b018ecc9d35 100644
> > > --- a/arch/arm64/kvm/debug.c
> > > +++ b/arch/arm64/kvm/debug.c
> > > @@ -233,7 +233,7 @@ void kvm_debug_handle_oslar(struct kvm_vcpu *vcpu, u64 val)
> > >    void kvm_enable_trbe(void)
> > >    {
> > >    	if (has_vhe() || is_protected_kvm_enabled() ||
> > > -	    WARN_ON_ONCE(preemptible()))
> > > +	    WARN_ON_ONCE(preemptible()) || !is_kvm_arm_initialised())
> >
> > Hi Yingchao,
> >
> > There shouldn't be a warning for this, at least for the case where
> > it's not initialized and never will be. If you're never going to run a
> > guest these functions can all skip, the same way for !has_vhe() etc.
>
> It's not a warning. It's a bona-fide crash:
>
> void kvm_enable_trbe(void)
> {
> 	if (has_vhe() || is_protected_kvm_enabled() ||
> 	    WARN_ON_ONCE(preemptible()))
> 		return;
>
> 	host_data_set_flag(TRBE_ENABLED); <--- Explodes here
> }
>
> So the write of the flag has to be skipped if KVM is available, even
> if KVM is compiled in.
>
> 	M.

Yeah. And just in case there is any confusion, I didn't mean that we 
should not have the check entirely, just that it shouldn't be in the 
WARN_ON_ONCE(). We should put it in the part that makes the functions 
silently skip:

  if (has_vhe() || is_protected_kvm_enabled() ||
      !is_kvm_arm_initialised() ||
      WARN_ON_ONCE(preemptible()))
          return;

And the same for kvm_disable_trbe() and kvm_tracing_set_el1_configuration().

James