Re: [PATCH v2 for-4.14 1/2] x86/mem_sharing: block interrupt injection for forks

[Tamas K Lengyel <tamas@xxxxxxxxxxxxx>]

On Mon, May 25, 2020 at 12:06 AM Jan Beulich <jbeulich@xxxxxxxx> wrote:
>
> On 22.05.2020 18:33, Tamas K Lengyel wrote:
> > When running shallow forks without device models it may be undesirable for 
> > Xen
> > to inject interrupts. With Windows forks we have observed the kernel going 
> > into
> > infinite loops when trying to process such interrupts, likely because it 
> > attempts
> > to interact with devices that are not responding without QEMU running. By
> > disabling interrupt injection the fuzzer can exercise the target code 
> > without
> > interference.
> >
> > Forks & memory sharing are only available on Intel CPUs so this only applies
> > to vmx.
>
> Looking at e.g. mem_sharing_control() I can't seem to be able to confirm
> this. Would you mind pointing me at where this restriction is coming from?

Both mem_access and mem_sharing are only implemented for EPT:
http://xenbits.xen.org/hg/xen-unstable.hg/file/5eadf9363c25/xen/arch/x86/mm/p2m-ept.c#l126.

>
> > --- a/xen/arch/x86/hvm/vmx/intr.c
> > +++ b/xen/arch/x86/hvm/vmx/intr.c
> > @@ -256,6 +256,12 @@ void vmx_intr_assist(void)
> >      if ( unlikely(v->arch.vm_event) && v->arch.vm_event->sync_event )
> >          return;
> >
> > +#ifdef CONFIG_MEM_SHARING
> > +    /* Block event injection for VM fork if requested */
> > +    if ( unlikely(v->domain->arch.hvm.mem_sharing.block_interrupts) )
> > +        return;
> > +#endif
>
> The two earlier returns are temporary as far as the guest is concerned,
> i.e. eventually the interrupt(s) will get delivered. The one you add
> looks as if it is a permanent thing, i.e. interrupt requests will pile
> up and potentially confuse a guest down the road. This _may_ be okay
> for your short-lived-shallow-fork scenario, but then wants at least
> calling out in the public header by a comment (and I think the same
> goes for XENMEM_FORK_WITH_IOMMU_ALLOWED that's already there).

This is indeed only for the short-lived forks, that's why this is an
optional flag that can be enabled when creating forks and it's not on
by default. In that use-case the VM executes for fractions of a second
and we want to only executes very specific code-segments with
absolutely no interference. Interrupts in that case are just a
nuisance that in the best case slow the fuzzing process down but as we
observed in the worst case complete stall it.

>
> > --- a/xen/include/asm-x86/hvm/domain.h
> > +++ b/xen/include/asm-x86/hvm/domain.h
> > @@ -74,6 +74,8 @@ struct mem_sharing_domain
> >       * to resume the search.
> >       */
> >      unsigned long next_shared_gfn_to_relinquish;
> > +
> > +    bool block_interrupts;
> >  };
>
> Please can you avoid unnecessary growth of the structure by inserting
> next to the pre-existing bool rather than at the end?

Sure. Do you want me to resend the patch for that?

Tamas