On Wed, Jan 08, 2020 at 06:46:30PM +0100, Paolo Bonzini wrote:
> On 23/12/19 21:10, Peter Xu wrote:
> >> Yes, kvm->slots_lock is taken by x86_set_memory_region. We need to move
> >> that to the callers, of which several are already taking the lock (all
> >> except vmx_set_tss_addr and kvm_arch_destroy_vm).
> > OK, will do. I'll directly replace the x86_set_memory_region() calls
> > in kvm_arch_destroy_vm() to be __x86_set_memory_region() since IIUC
> > the slots_lock is helpless when destroying the vm... then drop the
> > x86_set_memory_region() helper in the next version. Thanks,
>
> Be careful because it may cause issues with lockdep. Better just take
> the lock.
But you seemed to have fixed that already? :)
3898da947bba ("KVM: avoid using rcu_dereference_protected", 2017-08-02)
And this path is after kvm_destroy_vm() so kvm->users_count should be 0.
Or I feel like we need to have more places to take the lock..
--
Peter Xu
