Re: [PATCH v3 6/9] kvm: arm/arm64: Add host pmu to support VM introspection

[Christoffer Dall]

On Wed, Jan 18, 2017 at 06:05:46PM +0000, Mark Rutland wrote:
> On Wed, Jan 18, 2017 at 04:17:18PM +0000, Punit Agrawal wrote:
> > Mark Rutland <mark.rutland@xxxxxxx> writes:
> > 
> > > On Wed, Jan 18, 2017 at 02:51:31PM +0000, Punit Agrawal wrote:
> > >> I should've clarified in my reply that I wasn't looking to support the
> > >> third instance from Mark's examples above - "monitor all vCPUs on a
> > >> pCPU". I think it'll be quite expensive to figure out which threads from
> > >> a given pool are vCPUs.
> > >
> > > I'm not sure I follow why you would need to do that?
> > >
> > > In that case, we'd open a CPU-bound perf event for the pCPU, which would
> > > get installed in the CPU context immediately. It would be present for
> > > all tasks.
> > >
> > > Given it's present for all tasks, we don't need to figure out which
> > > happen to have vCPUs. The !vCPU tasks simply shouldn't trigger events.
> > >
> > > Am I missing something?
> > 
> > When enabling a CPU-bound event for pCPU, we'd have to enable trapping
> > of TLB operations for the vCPUs running on pCPU. Have a look at Patch
> > 7/9.
> > 
> > Also, we'd have to enable/disable trapping when tasks are migrated
> > between pCPUs.
> 
> Ah, so we can't configure the trap and leave it active, since it'll
> affect the host.
> 
> We could have a per-cpu flag, and a hook into vcpu_run, but that's also
> gnarly.
> 
> I'll have a think.
> 
> > So far I've assumed that a VM pid is immutable. If that doesn't hold
> > then we need to think of another mechanism to refer to a VM from
> > userspace.
> 
> Even if we can't migrate the VM between processes (i.e. it's immutable),
> it's still not unique within a process, so I'm fairly sure we need
> another mechanism (even if we get away with the common case today).
> 
I don't understand what the requirements here are exactly but the KVM
API documentation says:

  In general file descriptors can be migrated among processes by means
  of fork() and the SCM_RIGHTS facility of unix domain socket.  These
  kinds of tricks are explicitly not supported by kvm.  While they will
  not cause harm to the host, their actual behavior is not guaranteed by
  the API.  The only supported use is one virtual machine per process,
  and one vcpu per thread.

So this code should maintain those semantics and it's fair to assume the
thread group leader of a given VM stays the same, but the code must not
rely on this fact for safe operations.

I also don't see why a process couldn't open multiple VMs; however
messy that may be, it appears possible to me.

-Christoffer