Re: [PATCH 1/7] KVM: X86: Fix missed remote tlb flush in rmap_write_protect()

[Sean Christopherson <seanjc@google.com>]

On Tue, Aug 24, 2021, Lai Jiangshan wrote:
> From: Lai Jiangshan <laijs@linux.alibaba.com>
> 
> When kvm->tlbs_dirty > 0, some rmaps might have been deleted
> without flushing tlb remotely after kvm_sync_page().  If @gfn
> was writable before and it's rmaps was deleted in kvm_sync_page(),
> we need to flush tlb too even if __rmap_write_protect() doesn't
> request it.
> 
> Fixes: 4731d4c7a077 ("KVM: MMU: out of sync shadow core")

Should be

Fixes: a4ee1ca4a36e ("KVM: MMU: delay flush all tlbs on sync_page path")

> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
> ---
>  arch/x86/kvm/mmu/mmu.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
> 
> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> index 4853c033e6ce..313918df1a10 100644
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -1420,6 +1420,14 @@ bool kvm_mmu_slot_gfn_write_protect(struct kvm *kvm,
>  			rmap_head = gfn_to_rmap(gfn, i, slot);
>  			write_protected |= __rmap_write_protect(kvm, rmap_head, true);
>  		}
> +		/*
> +		 * When kvm->tlbs_dirty > 0, some rmaps might have been deleted
> +		 * without flushing tlb remotely after kvm_sync_page().  If @gfn
> +		 * was writable before and it's rmaps was deleted in kvm_sync_page(),
> +		 * we need to flush tlb too.
> +		 */
> +		if (min_level == PG_LEVEL_4K && kvm->tlbs_dirty)
> +			write_protected = true;
>  	}
>  
>  	if (is_tdp_mmu_enabled(kvm))
> @@ -5733,6 +5741,14 @@ void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
>  		flush = slot_handle_level(kvm, memslot, slot_rmap_write_protect,
>  					  start_level, KVM_MAX_HUGEPAGE_LEVEL,
>  					  false);
> +		/*
> +		 * When kvm->tlbs_dirty > 0, some rmaps might have been deleted
> +		 * without flushing tlb remotely after kvm_sync_page().  If @gfn
> +		 * was writable before and it's rmaps was deleted in kvm_sync_page(),
> +		 * we need to flush tlb too.
> +		 */
> +		if (start_level == PG_LEVEL_4K && kvm->tlbs_dirty)
> +			flush = true;
>  		write_unlock(&kvm->mmu_lock);
>  	}

My vote is to do a revert of a4ee1ca4a36e with slightly less awful batching, and
then improve the batching even further if there's a noticeable loss of performance
(or just tell people to stop using shadow paging :-D).  Zapping SPTEs but not
flushing is just asking for these types of whack-a-mole bugs.

E.g. instead of a straight revert, do this for sync_page():

diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h
index 50ade6450ace..1fca27a08c00 100644
--- a/arch/x86/kvm/mmu/paging_tmpl.h
+++ b/arch/x86/kvm/mmu/paging_tmpl.h
@@ -1095,13 +1095,7 @@ static int FNAME(sync_page)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
                        return 0;

                if (FNAME(prefetch_invalid_gpte)(vcpu, sp, &sp->spt[i], gpte)) {
-                       /*
-                        * Update spte before increasing tlbs_dirty to make
-                        * sure no tlb flush is lost after spte is zapped; see
-                        * the comments in kvm_flush_remote_tlbs().
-                        */
-                       smp_wmb();
-                       vcpu->kvm->tlbs_dirty++;
+                       set_spte_ret |= SET_SPTE_NEED_REMOTE_TLB_FLUSH;
                        continue;
                }

@@ -1116,12 +1110,7 @@ static int FNAME(sync_page)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)

                if (gfn != sp->gfns[i]) {
                        drop_spte(vcpu->kvm, &sp->spt[i]);
-                       /*
-                        * The same as above where we are doing
-                        * prefetch_invalid_gpte().
-                        */
-                       smp_wmb();
-                       vcpu->kvm->tlbs_dirty++;
+                       set_spte_ret |= SET_SPTE_NEED_REMOTE_TLB_FLUSH;
                        continue;
                }