Re: [PATCH v2 2/5] KVM: MMU: Convert remote flushes to kvm_mark_tlb_dirty() and a conditional flush

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, May 17, 2012 at 01:24:41PM +0300, Avi Kivity wrote:
> This allows us to later move the actual flush out of protection of the
> mmu spinlock, provided there are no additional dependencies.
> 
> Constructs of the form
> 
>     if (pred)
>         kvm_flush_remote_tlbs(kvm)
> 
> are converted to
> 
>     if (pred)
>         kvm_mark_tlb_dirty(kvm)
> 
>     kvm_cond_flush_remote_tlbs(kvm)
> 
> so that if another thread caused pred to transition from true to false,
> but has not yet flushed the tlb, we notice it and flush it before proceeding.
> 
> Signed-off-by: Avi Kivity <avi@redhat.com>
> ---
>  arch/ia64/kvm/kvm-ia64.c   |    6 ++++--
>  arch/x86/kvm/mmu.c         |   45 +++++++++++++++++++++++++++-----------------
>  arch/x86/kvm/paging_tmpl.h |    4 +++-
>  arch/x86/kvm/x86.c         |    4 +++-
>  virt/kvm/kvm_main.c        |    4 +++-
>  5 files changed, 41 insertions(+), 22 deletions(-)
> 
> diff --git a/arch/ia64/kvm/kvm-ia64.c b/arch/ia64/kvm/kvm-ia64.c
> index bd77cb5..a4a92d8 100644
> --- a/arch/ia64/kvm/kvm-ia64.c
> +++ b/arch/ia64/kvm/kvm-ia64.c
> @@ -1628,7 +1628,8 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
>  
>  void kvm_arch_flush_shadow(struct kvm *kvm)
>  {
> -	kvm_flush_remote_tlbs(kvm);
> +	kvm_mark_tlb_dirty(kvm);
> +	kvm_cond_flush_remote_tlbs(kvm);
>  }
>  
>  long kvm_arch_dev_ioctl(struct file *filp,
> @@ -1853,10 +1854,11 @@ int kvm_vm_ioctl_get_dirty_log(struct kvm *kvm,
>  
>  	/* If nothing is dirty, don't bother messing with page tables. */
>  	if (is_dirty) {
> -		kvm_flush_remote_tlbs(kvm);
> +		kvm_mark_tlb_dirty(kvm);
>  		n = kvm_dirty_bitmap_bytes(memslot);
>  		memset(memslot->dirty_bitmap, 0, n);
>  	}
> +	kvm_cond_flush_remote_tlbs(kvm);
>  	r = 0;
>  out:
>  	mutex_unlock(&kvm->slots_lock);
> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> index 07424cf..03a9c80 100644
> --- a/arch/x86/kvm/mmu.c
> +++ b/arch/x86/kvm/mmu.c
> @@ -1170,7 +1170,9 @@ static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
>  	}
>  
>  	if (need_flush)
> -		kvm_flush_remote_tlbs(kvm);
> +		kvm_mark_tlb_dirty(kvm);
> +
> +	kvm_cond_flush_remote_tlbs(kvm);
>  
>  	return 0;
>  }
> @@ -1294,7 +1296,8 @@ static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
>  	rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
>  
>  	kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
> -	kvm_flush_remote_tlbs(vcpu->kvm);
> +	kvm_mark_tlb_dirty(vcpu->kvm);
> +	kvm_cond_flush_remote_tlbs(vcpu->kvm);
>  }
>  
>  int kvm_age_hva(struct kvm *kvm, unsigned long hva)
> @@ -1697,7 +1700,9 @@ static void mmu_sync_children(struct kvm_vcpu *vcpu,
>  			protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
>  
>  		if (protected)
> -			kvm_flush_remote_tlbs(vcpu->kvm);
> +			kvm_mark_tlb_dirty(vcpu->kvm);
> +
> +		kvm_cond_flush_remote_tlbs(vcpu->kvm);
>  
>  		for_each_sp(pages, sp, parents, i) {
>  			kvm_sync_page(vcpu, sp, &invalid_list);
> @@ -1786,7 +1791,8 @@ static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
>  		&vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
>  	if (!direct) {
>  		if (rmap_write_protect(vcpu->kvm, gfn))
> -			kvm_flush_remote_tlbs(vcpu->kvm);
> +			kvm_mark_tlb_dirty(vcpu->kvm);
> +		kvm_cond_flush_remote_tlbs(vcpu->kvm);
>  		if (level > PT_PAGE_TABLE_LEVEL && need_sync)
>  			kvm_sync_pages(vcpu, gfn);
>  
> @@ -1861,7 +1867,8 @@ static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
>  	if (is_large_pte(*sptep)) {
>  		drop_spte(vcpu->kvm, sptep);
>  		--vcpu->kvm->stat.lpages;
> -		kvm_flush_remote_tlbs(vcpu->kvm);
> +		kvm_mark_tlb_dirty(vcpu->kvm);
> +		kvm_cond_flush_remote_tlbs(vcpu->kvm);
>  	}
>  }
>  
> @@ -1883,7 +1890,8 @@ static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
>  			return;
>  
>  		drop_parent_pte(child, sptep);
> -		kvm_flush_remote_tlbs(vcpu->kvm);
> +		kvm_mark_tlb_dirty(vcpu->kvm);
> +		kvm_cond_flush_remote_tlbs(vcpu->kvm);
>  	}
>  }
>  
> @@ -2021,7 +2029,8 @@ static void kvm_mmu_commit_zap_page(struct kvm *kvm,
>  	if (list_empty(invalid_list))
>  		return;
>  
> -	kvm_flush_remote_tlbs(kvm);
> +	kvm_mark_tlb_dirty(kvm);
> +	kvm_cond_flush_remote_tlbs(kvm);
>  
>  	if (atomic_read(&kvm->arch.reader_counter)) {
>  		kvm_mmu_isolate_pages(invalid_list);
> @@ -2344,7 +2353,9 @@ static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
>  	 * might be cached on a CPU's TLB.
>  	 */
>  	if (is_writable_pte(entry) && !is_writable_pte(*sptep))
> -		kvm_flush_remote_tlbs(vcpu->kvm);
> +		kvm_mark_tlb_dirty(vcpu->kvm);
> +
> +	kvm_cond_flush_remote_tlbs(vcpu->kvm);
>  done:
>  	return ret;
>  }
> @@ -2376,15 +2387,15 @@ static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
>  
>  			child = page_header(pte & PT64_BASE_ADDR_MASK);
>  			drop_parent_pte(child, sptep);
> -			kvm_flush_remote_tlbs(vcpu->kvm);
>  		} else if (pfn != spte_to_pfn(*sptep)) {
>  			pgprintk("hfn old %llx new %llx\n",
>  				 spte_to_pfn(*sptep), pfn);
>  			drop_spte(vcpu->kvm, sptep);
> -			kvm_flush_remote_tlbs(vcpu->kvm);
> +			kvm_mark_tlb_dirty(vcpu->kvm);
>  		} else
>  			was_rmapped = 1;
>  	}
> +	kvm_cond_flush_remote_tlbs(vcpu->kvm);
>  
>  	if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
>  		      level, gfn, pfn, speculative, true,
> @@ -3561,14 +3572,13 @@ static bool need_remote_flush(u64 old, u64 new)
>  }
>  
>  static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
> -				    bool remote_flush, bool local_flush)
> +				    bool local_flush)
>  {
>  	if (zap_page)
>  		return;
>  
> -	if (remote_flush)
> -		kvm_flush_remote_tlbs(vcpu->kvm);
> -	else if (local_flush)
> +	kvm_cond_flush_remote_tlbs(vcpu->kvm);
> +	if (local_flush)
>  		kvm_mmu_flush_tlb(vcpu);
>  }
>  
> @@ -3742,11 +3752,11 @@ void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
>  			      & mask.word) && rmap_can_add(vcpu))
>  				mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
>  			if (!remote_flush && need_remote_flush(entry, *spte))
> -				remote_flush = true;
> +				kvm_mark_tlb_dirty(vcpu->kvm);


A remote cpu can flush the TLBs, marking the TLB as synced. 
But "remote_flush" variable remains true, so for further spte updates,
the TLB is not marked dirty.

Better kill the remote_flush variable (still reviewing).