Re: [PATCH 05/15] perf: Track guest callbacks on a per-CPU basis

[Sean Christopherson <seanjc@google.com>]

On Fri, Aug 27, 2021, Peter Zijlstra wrote:
> On Fri, Aug 27, 2021 at 02:49:50PM +0000, Sean Christopherson wrote:
> > On Fri, Aug 27, 2021, Peter Zijlstra wrote:
> > > On Thu, Aug 26, 2021 at 05:57:08PM -0700, Sean Christopherson wrote:
> > > > Use a per-CPU pointer to track perf's guest callbacks so that KVM can set
> > > > the callbacks more precisely and avoid a lurking NULL pointer dereference.
> > > 
> > > I'm completely failing to see how per-cpu helps anything here...
> > 
> > It doesn't help until KVM is converted to set the per-cpu pointer in flows that
> > are protected against preemption, and more specifically when KVM only writes to
> > the pointer from the owning CPU.  
> 
> So the 'problem' I have with this is that sane (!KVM using) people, will
> still have to suffer that load, whereas with the static_call() we patch
> in an 'xor %rax,%rax' and only have immediate code flow.

Again, I've no objection to the static_call() approach.  I didn't even see the
patch until I had finished testing my series :-/

> > Ignoring static call for the moment, I don't see how the unreg side can be safe
> > using a bare single global pointer.  There is no way for KVM to prevent an NMI
> > from running in parallel on a different CPU.  If there's a more elegant solution,
> > especially something that can be backported, e.g. an rcu-protected pointer, I'm
> > all for it.  I went down the per-cpu path because it allowed for cleanups in KVM,
> > but similar cleanups can be done without per-cpu perf callbacks.
> 
> If all the perf_guest_cbs dereferences are with preemption disabled
> (IRQs disabled, IRQ context, NMI context included), then the sequence:
> 
> 	WRITE_ONCE(perf_guest_cbs, NULL);
> 	synchronize_rcu();
> 
> Ensures that all prior observers of perf_guest_csb will have completed
> and future observes must observe the NULL value.

That alone won't be sufficient, as the read side also needs to ensure it doesn't
reload perf_guest_cbs between NULL checks and dereferences.  But that's easy
enough to solve with a READ_ONCE and maybe a helper to make it more cumbersome
to use perf_guest_cbs directly.

How about this for a series?

  1. Use READ_ONCE/WRITE_ONCE + synchronize_rcu() to fix the underlying bug
  2. Fix KVM PT interrupt handler bug
  3. Kill off perf_guest_cbs usage in architectures that don't need the callbacks
  4. Replace ->is_in_guest()/->is_user_mode() with ->state(), and s/get_guest_ip/get_ip
  5. Implement static_call() support
  6. Cleanups, if there are any
  6..N KVM cleanups, e.g. to eliminate current_vcpu and share x86+arm64 callbacks