Re: [patch] mm, oom: fix unnecessary killing of additional processes

[Michal Hocko <mhocko@kernel.org>]

On Thu 21-06-18 13:50:53, David Rientjes wrote:
> On Thu, 21 Jun 2018, Michal Hocko wrote:
> 
> > > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > > index 6bcecc325e7e..ac08f5d711be 100644
> > > > --- a/arch/x86/kvm/x86.c
> > > > +++ b/arch/x86/kvm/x86.c
> > > > @@ -7203,8 +7203,9 @@ static void vcpu_load_eoi_exitmap(struct kvm_vcpu *vcpu)
> > > >  	kvm_x86_ops->load_eoi_exitmap(vcpu, eoi_exit_bitmap);
> > > >  }
> > > >  
> > > > -void kvm_arch_mmu_notifier_invalidate_range(struct kvm *kvm,
> > > > -		unsigned long start, unsigned long end)
> > > > +int kvm_arch_mmu_notifier_invalidate_range(struct kvm *kvm,
> > > > +		unsigned long start, unsigned long end,
> > > > +		bool blockable)
> > > >  {
> > > >  	unsigned long apic_address;
> > > >  
> > > > @@ -7215,6 +7216,8 @@ void kvm_arch_mmu_notifier_invalidate_range(struct kvm *kvm,
> > > >  	apic_address = gfn_to_hva(kvm, APIC_DEFAULT_PHYS_BASE >> PAGE_SHIFT);
> > > >  	if (start <= apic_address && apic_address < end)
> > > >  		kvm_make_all_cpus_request(kvm, KVM_REQ_APIC_PAGE_RELOAD);
> > > > +
> > > > +	return 0;
> > > >  }
> > > >  
> > > >  void kvm_vcpu_reload_apic_access_page(struct kvm_vcpu *vcpu)
> > > 
> > > Auditing the first change in the patch, this is incorrect because 
> > > kvm_make_all_cpus_request() for KVM_REQ_APIC_PAGE_RELOAD can block in 
> > > kvm_kick_many_cpus() and that is after kvm_make_request() has been done.
> > 
> > I would have to check the code closer. But doesn't
> > kvm_make_all_cpus_request call get_cpu which is preempt_disable? I
> > definitely plan to talk to respective maintainers about these changes of
> > course.
> > 
> 
> preempt_disable() is required because it calls kvm_kick_many_cpus() with 
> wait == true because KVM_REQ_APIC_PAGE_RELOAD sets KVM_REQUEST_WAIT and 
> thus the smp_call_function_many() is going to block until all cpus can run 
> ack_flush().

I will make sure to talk to the maintainer of the respective code to
do the nonblock case correctly.

-- 
Michal Hocko
SUSE Labs