Re: [Xen-devel] [PATCH 5/8] x86/vm-event/monitor: don't compromise monitor_write_data on domain cleanup

[Corneliu ZUZU <czuzu@xxxxxxxxxxxxxxx>]

On 7/4/2016 4:50 PM, Razvan Cojocaru wrote:
> On 07/04/16 16:11, Jan Beulich wrote:
> > > > > On 04.07.16 at 15:03, <czuzu@xxxxxxxxxxxxxxx> wrote:
> > > On 7/4/2016 3:47 PM, Jan Beulich wrote:
> > > > > > > On 30.06.16 at 20:45, <czuzu@xxxxxxxxxxxxxxx> wrote:
> > > > > The arch_vm_event structure is dynamically allocated and freed @
> > > > > vm_event_cleanup_domain. This cleanup is triggered e.g. when the toolstack user
> > > > > disables domain monitoring (xc_monitor_disable), which in turn effectively
> > > > > discards any information that was in arch_vm_event.write_data.
> > > > Isn't that rather a toolstack user bug, not warranting a relatively
> > > > extensive (even if mostly mechanical) hypervisor change like this
> > > > one? Sane monitor behavior, after all, is required anyway for the
> > > > monitored guest to survive.
> > > Sorry but could you please rephrase this, I don't quite understand what
> > > you're saying.
> > > The write_data field in arch_vm_event should _not ever_ be invalidated
> > > as a direct result of a toolstack user's action.
> > The monitoring app can cause all kinds of problems to the guest it
> > monitors. Why would this specific one need taking care of in the
> > hypervisor, instead of demanding that the app not disable monitoring
> > at the wrong time?
> I'm not sure there's a right time here. The problem is that, if I
> understand this correctly, a race is possible between the moment the
> userspace application responds to the vm_event _and_ call
> xc_monitor_disable() and the time hvm_do_resume() gets called.
>
> If xc_monitor_disable() happened before hvm_do_resume() springs into
> action, we lose a register write. There's no guaranteed way for this not
> to happen as far as I can see, although it's true that the race should
> pretty much never happen in practice - at least we've never come across
> such a case so far.
>
>
> Thanks,
> Razvan
Perfectly pointed, thanks. Note that xc_monitor_disable() may happen 
before hvm_do_resume() because the latter only happens _when the 
scheduler reschedules the target vCPU_, which _may not happen between_ 
the moment the toolstack user _handles the vm-event_ and the moment when 
he _calls xc_monitor_disable()_, but rather after xc_monitor_disable() 
is called.
Corneliu.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel