Re: [PATCH v2 for-4.14 3/3] xen/vm_event: Add safe to disable vm_event

[Tamas K Lengyel <tamas@tklengyel.com>]

On Tue, Jun 2, 2020 at 6:54 AM Roger Pau Monné <roger.pau@citrix.com> wrote:
>
> On Wed, May 20, 2020 at 08:31:54PM -0600, Tamas K Lengyel wrote:
> > Instead of having to repeatedly try to disable vm_events,
>
> Why not use a hypercall continuation instead so that this is all
> hidden from the caller?
>
> I take that the current interface requires the user to repeatedly
> issue hypercalls in order to disable vm_events until one of those
> succeeds?

No, it succeeds right away. And then the guest crashes in unique and
unpredictable ways.

>
> > request a specific
> > vm_event to be sent when the domain is safe to continue with shutting down
> > the vm_event interface.
> >
> > Signed-off-by: Tamas K Lengyel <tamas@tklengyel.com>
> > ---
> >  xen/arch/x86/hvm/hvm.c            | 38 ++++++++++++++++++++++++++-----
> >  xen/arch/x86/hvm/monitor.c        | 14 ++++++++++++
> >  xen/arch/x86/monitor.c            | 13 +++++++++++
> >  xen/include/asm-x86/domain.h      |  1 +
> >  xen/include/asm-x86/hvm/monitor.h |  1 +
> >  xen/include/public/domctl.h       |  2 ++
> >  xen/include/public/vm_event.h     |  8 +++++++
> >  7 files changed, 71 insertions(+), 6 deletions(-)
> >
> > diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
> > index e6780c685b..fc7e1e2b22 100644
> > --- a/xen/arch/x86/hvm/hvm.c
> > +++ b/xen/arch/x86/hvm/hvm.c
> > @@ -563,15 +563,41 @@ void hvm_do_resume(struct vcpu *v)
> >          v->arch.hvm.inject_event.vector = HVM_EVENT_VECTOR_UNSET;
> >      }
> >
> > -    if ( unlikely(v->arch.vm_event) && v->arch.monitor.next_interrupt_enabled )
> > +    if ( unlikely(v->arch.vm_event) )
> >      {
> > -        struct x86_event info;
> > +        struct domain *d = v->domain;
> > +
> > +        if ( v->arch.monitor.next_interrupt_enabled )
> > +        {
> > +            struct x86_event info;
> > +
> > +            if ( hvm_get_pending_event(v, &info) )
> > +            {
> > +                hvm_monitor_interrupt(info.vector, info.type, info.error_code,
> > +                                      info.cr2);
> > +                v->arch.monitor.next_interrupt_enabled = false;
> > +            }
> > +        }
> >
> > -        if ( hvm_get_pending_event(v, &info) )
> > +        if ( d->arch.monitor.safe_to_disable )
> >          {
> > -            hvm_monitor_interrupt(info.vector, info.type, info.error_code,
> > -                                  info.cr2);
> > -            v->arch.monitor.next_interrupt_enabled = false;
> > +            const struct vcpu *check_vcpu;
> > +            bool pending_op = false;
> > +
> > +            for_each_vcpu ( d, check_vcpu )
> > +            {
> > +                if ( vm_event_check_pending_op(check_vcpu) )
>
> Don't you need some kind of lock here, since you are poking at another
> vCPU which could be modifying any of those bits?
>
> > +                {
> > +                    pending_op = true;
> > +                    break;
> > +                }
> > +            }
> > +
> > +            if ( !pending_op )
> > +            {
> > +                hvm_monitor_safe_to_disable();
> > +                d->arch.monitor.safe_to_disable = false;
> > +            }
> >          }
> >      }
> >  }
> > diff --git a/xen/arch/x86/hvm/monitor.c b/xen/arch/x86/hvm/monitor.c
> > index f5d89e71d1..75fd1a4b68 100644
> > --- a/xen/arch/x86/hvm/monitor.c
> > +++ b/xen/arch/x86/hvm/monitor.c
> > @@ -300,6 +300,20 @@ bool hvm_monitor_check_p2m(unsigned long gla, gfn_t gfn, uint32_t pfec,
> >      return monitor_traps(curr, true, &req) >= 0;
> >  }
> >
> > +void hvm_monitor_safe_to_disable(void)
> > +{
> > +    struct vcpu *curr = current;
> > +    struct arch_domain *ad = &curr->domain->arch;
>
> const
>
> > +    vm_event_request_t req = {};
> > +
> > +    if ( !ad->monitor.safe_to_disable )
> > +        return;
>
> Should this rather be an ASSERT? I don't think you are supposed to
> call hvm_monitor_safe_to_disable when the bit is not set?
>
> > +
> > +    req.reason = VM_EVENT_REASON_SAFE_TO_DISABLE;
>
> I think you cat set the field at definition time.
>
> > +
> > +    monitor_traps(curr, 0, &req);
> > +}
> > +
> >  /*
> >   * Local variables:
> >   * mode: C
> > diff --git a/xen/arch/x86/monitor.c b/xen/arch/x86/monitor.c
> > index 1517a97f50..86e0ba2fbc 100644
> > --- a/xen/arch/x86/monitor.c
> > +++ b/xen/arch/x86/monitor.c
> > @@ -339,6 +339,19 @@ int arch_monitor_domctl_event(struct domain *d,
> >          break;
> >      }
> >
> > +    case XEN_DOMCTL_MONITOR_EVENT_SAFE_TO_DISABLE:
> > +    {
> > +        bool old_status = ad->monitor.safe_to_disable;
> > +
> > +        if ( unlikely(old_status == requested_status) )
> > +            return -EEXIST;
> > +
> > +        domain_pause(d);
> > +        ad->monitor.safe_to_disable = requested_status;
>
> Maybe I'm missing something, but I don't see any check that others
> events are disabled before safe_to_disable is set?
>
> In the same way, you should prevent setting any events when
> safe_to_disable is set IMO, likely returning -EBUSY in both cases.
>
> Thanks, Roger.

Thanks for the feedback again. I won't have the bandwidth to address
these so I'm dropping this patch. If Bitdefender is so inclined to
pick-up later they are welcome to do so. This is only needed if their
buggy feature is enabled.

Tamas