Re: [PATCH v3 6/6] KVM: MMU: delay flush all tlbs on sync_page path

[Marcelo Tosatti <mtosatti@redhat.com>]

On Mon, Nov 22, 2010 at 11:45:18AM +0800, Xiao Guangrong wrote:
> On 11/20/2010 12:11 AM, Marcelo Tosatti wrote:
> 
> >>  void kvm_flush_remote_tlbs(struct kvm *kvm)
> >>  {
> >> +	int dirty_count = atomic_read(&kvm->tlbs_dirty);
> >> +
> >> +	smp_mb();
> >>  	if (make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
> >>  		++kvm->stat.remote_tlb_flush;
> >> +	atomic_sub(dirty_count, &kvm->tlbs_dirty);
> >>  }
> > 
> > This is racy because kvm_flush_remote_tlbs might be called without
> > mmu_lock protection.
> 
> Sorry for my carelessness, it should be 'cmpxchg' here.
> 
> > You could decrease the counter on
> > invalidate_page/invalidate_range_start only, 
> 
> I want to avoid a unnecessary tlbs flush, if tlbs have been flushed
> after sync_page, then we don't need flush tlbs on invalidate_page/
> invalidate_range_start path.
> 
> > these are not fast paths
> > anyway.
> > 
> 
> How about below patch? it just needs one atomic operation.
> 
> ---
>  arch/x86/kvm/paging_tmpl.h |    4 ++--
>  include/linux/kvm_host.h   |    2 ++
>  virt/kvm/kvm_main.c        |    7 ++++++-
>  3 files changed, 10 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/x86/kvm/paging_tmpl.h b/arch/x86/kvm/paging_tmpl.h
> index dfb906f..e64192f 100644
> --- a/arch/x86/kvm/paging_tmpl.h
> +++ b/arch/x86/kvm/paging_tmpl.h
> @@ -781,14 +781,14 @@ static int FNAME(sync_page)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
>  		gfn = gpte_to_gfn(gpte);
>  
>  		if (FNAME(map_invalid_gpte)(vcpu, sp, &sp->spt[i], gpte)) {
> -			kvm_flush_remote_tlbs(vcpu->kvm);
> +			vcpu->kvm->tlbs_dirty++;
>  			continue;
>  		}
>  
>  		if (gfn != sp->gfns[i]) {
>  			drop_spte(vcpu->kvm, &sp->spt[i],
>  				      shadow_trap_nonpresent_pte);
> -			kvm_flush_remote_tlbs(vcpu->kvm);
> +			vcpu->kvm->tlbs_dirty++;
>  			continue;
>  		}
>  
> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> index 4bd663d..dafd90e 100644
> --- a/include/linux/kvm_host.h
> +++ b/include/linux/kvm_host.h
> @@ -249,6 +249,7 @@ struct kvm {
>  	struct mmu_notifier mmu_notifier;
>  	unsigned long mmu_notifier_seq;
>  	long mmu_notifier_count;
> +	long tlbs_dirty;
>  #endif
>  };
>  
> @@ -377,6 +378,7 @@ void kvm_vcpu_on_spin(struct kvm_vcpu *vcpu);
>  void kvm_resched(struct kvm_vcpu *vcpu);
>  void kvm_load_guest_fpu(struct kvm_vcpu *vcpu);
>  void kvm_put_guest_fpu(struct kvm_vcpu *vcpu);
> +
>  void kvm_flush_remote_tlbs(struct kvm *kvm);
>  void kvm_reload_remote_mmus(struct kvm *kvm);
>  
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index fb93ff9..fe0a1a7 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -168,8 +168,12 @@ static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
>  
>  void kvm_flush_remote_tlbs(struct kvm *kvm)
>  {
> +	long dirty_count = kvm->tlbs_dirty;
> +
> +	smp_mb();
>  	if (make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
>  		++kvm->stat.remote_tlb_flush;

<---

> +	cmpxchg(&kvm->tlbs_dirty, dirty_count, 0);
>  }


Still problematic, if tlbs_dirty is set in the point indicated above.

Invalidate page should be quite rare, so checking for tlb_dirty only
there is OK.