Re: [RFC PATCH 0/2] Expose available KVM free memory slot count to help avoid aborts

[Avi Kivity <avi@redhat.com>]

On 02/23/2011 11:46 PM, Alex Williamson wrote:
> >
> >  But kvm_arch_flush_shadow() takes mmu_lock currently, so that needs
> >  fixing.
>
> Hmm, I tried to follow the example in the !npages path just above this
> that does:
>
> rcu_assign_pointer()
> synchronize_srcu_expedited()
> kvm_arch_flush_shadow()
>
> Do we have an existing issue there with mmu_lock?  Thanks,
>

It's the issue I pointed out - after rcu_assign_pointer and before 
kvm_arch_flush_shadow() we can still get shadow pages created, so we 
have a mix of direct and indirect bitmap.  Either my solution (recording 
whether the bitmap is direct or not in kvm_mmu_page) or Marcelo's 
(preventing shadow instantiation during this period) should work.

Hmm, your patch dereferences kvm->memslots without rcu_dereference().  
That's a mortal (as in oops) sin.  Note that sticking an 
rcu_dereference() blindly may or may not work - we might end up using 
information from two different generations of kvm->memslots, and using 
inconsistent data.

-- 
error compiling committee.c: too many arguments to function