KVM: Defer remote tlb flushes on invlpg (v2)

KVM flushes tlbs on remote cpus for two purposes: to protect guest pages that it needs to collect information about, and to prevent stale tlb entries from pointing to pages that no longer belong to the guest.

We can defer the latter flushes to the point when we actually free the pages, which is during an mmu notifier invocation. To this end, we add a new state remote_tlbs_dirty which marks whether the guest tlb might be inconsistent with the the shadow page tables. Whenever we do a conditional flush of remote tlbs, we check this state, and if the remote tlbs are dirty we flush them to ensure no inconsistency.

[v2: add help kvm_flush_remote_tlbs_cond() to remove the need for callers
to care about the new logic]

KVM flushes tlbs on remote cpus for two purposes: to protect guest pages that it needs to collect information about, and to prevent stale tlb entries from pointing to pages that no longer belong to the guest.

Quoted 4 lines

KVM flushes tlbs on remote cpus for two purposes: to protect guest pages that it needs to collect information about, and to prevent stale tlb entries from pointing to pages that no longer belong to the guest.

I assume you have numbers with some benchmark how much that improves performance. It would be nice to add them to the commit log.

KVM flushes tlbs on remote cpus for two purposes: to protect guest pages that it needs to collect information about, and to prevent stale tlb entries from pointing to pages that no longer belong to the guest.

We can defer the latter flushes to the point when we actually free the pages, which is during an mmu notifier invocation. To this end, we add a new state remote_tlbs_dirty which marks whether the guest tlb might be inconsistent with the the shadow page tables. Whenever we do a conditional flush of remote tlbs, we check this state, and if the remote tlbs are dirty we flush them to ensure no inconsistency.

[v2: add help kvm_flush_remote_tlbs_cond() to remove the need for callers
to care about the new logic]

arch/x86/kvm/paging_tmpl.h +0 -0

index 855eb71..18abdf9 100644
@@ -475,8 +475,10 @@ static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva)
 			break;
 	}

virt/kvm/kvm_main.c +14 -1

index 68b217e..2ee6a6d 100644
@@ -758,10 +758,22 @@ static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
 
 void kvm_flush_remote_tlbs(struct kvm *kvm)
 {
+	kvm->remote_tlbs_dirty = false;
+	wmb();
 	if (make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
 		++kvm->stat.remote_tlb_flush;
 }
 
+void kvm_flush_remote_tlbs_cond(struct kvm *kvm, bool cond)
+{
+	if (!cond) {
+		rmb();
+		cond = kvm->remote_tlbs_dirty;
+	}
+	if (cond)
+		kvm_flush_remote_tlbs(kvm);
+}
+
 void kvm_reload_remote_mmus(struct kvm *kvm)
 {
 	make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD);

@@ -840,8 +852,9 @@ static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
 	need_tlb_flush = kvm_unmap_hva(kvm, address);
 	spin_unlock(&kvm->mmu_lock);
 
+	rmb();
 	/* we've to flush the tlb before the pages can be freed */
-	if (need_tlb_flush)
+	if (need_tlb_flush || kvm->remote_tlbs_dirty)
 		kvm_flush_remote_tlbs(kvm);
 
 }

Quoted 9 lines

static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp) {

arch/x86/kvm/mmu.c +2 -0

+	bool need_flush;
+
 	if (sp->role.glevels != vcpu->arch.mmu.root_level) {
 		kvm_mmu_zap_page(vcpu->kvm, sp);
 		return 1;
 	}

arch/x86/kvm/mmu.c static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp); · line 1104

Quoted 5 lines

arch/x86/kvm/mmu.c +2 -2

-	if (rmap_write_protect(vcpu->kvm, sp->gfn))
-		kvm_flush_remote_tlbs(vcpu->kvm);
+	need_flush = rmap_write_protect(vcpu->kvm, sp->gfn);
+	kvm_flush_remote_tlbs_cond(vcpu->kvm, need_flush);

Why exactly do we need to flush remote and local tlbs here (in the cr3 overwrite path which is a local flush) if no change happened to sptes related to the current process? How is it relevant that a previous invlpg has run and we did only a local flush at the time invlpg run?

Same here.

Same here.

Quoted 12 lines

arch/x86/kvm/paging_tmpl.h +4 -2

@@ -475,8 +475,10 @@ static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva)
 			break;
 	}
 
-	if (need_flush)
-		kvm_flush_remote_tlbs(vcpu->kvm);
+	if (need_flush) {
+		vcpu->kvm->remote_tlbs_dirty = true;
+		kvm_x86_ops->tlb_flush(vcpu);
+	}
 	spin_unlock(&vcpu->kvm->mmu_lock);
 

arch/x86/kvm/paging_tmpl.h static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva) · line 484

Quoted 2 lines

arch/x86/kvm/paging_tmpl.h +0 -0

 	if (pte_gpa == -1)

I think it'd be more consistent to set_bit(KVM_REQ_TLB_FLUSH, &vcpu->requests) instead of invoking tlb_flush from a different place.

Quoted 35 lines

virt/kvm/kvm_main.c +14 -1

@@ -758,10 +758,22 @@ static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
 
 void kvm_flush_remote_tlbs(struct kvm *kvm)
 {
+	kvm->remote_tlbs_dirty = false;
+	wmb();
 	if (make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
 		++kvm->stat.remote_tlb_flush;
 }
 
+void kvm_flush_remote_tlbs_cond(struct kvm *kvm, bool cond)
+{
+	if (!cond) {
+		rmb();
+		cond = kvm->remote_tlbs_dirty;
+	}
+	if (cond)
+		kvm_flush_remote_tlbs(kvm);
+}
+
 void kvm_reload_remote_mmus(struct kvm *kvm)
 {
 	make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD);

@@ -840,8 +852,9 @@ static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
 	need_tlb_flush = kvm_unmap_hva(kvm, address);
 	spin_unlock(&kvm->mmu_lock);
 
+	rmb();
 	/* we've to flush the tlb before the pages can be freed */
-	if (need_tlb_flush)
+	if (need_tlb_flush || kvm->remote_tlbs_dirty)
 		kvm_flush_remote_tlbs(kvm);
 
 }

virt/kvm/kvm_main.c static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, · line 886

Quoted 10 lines

virt/kvm/kvm_main.c +2 -1

@@ -865,8 +878,9 @@ static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
 		need_tlb_flush |= kvm_unmap_hva(kvm, start);
 	spin_unlock(&kvm->mmu_lock);
 
+	rmb();
 	/* we've to flush the tlb before the pages can be freed */
-	if (need_tlb_flush)
+	if (need_tlb_flush || kvm->remote_tlbs_dirty)
 		kvm_flush_remote_tlbs(kvm);
 }
 

I think the only memory barrier required is a smp_mb() between setting remote_tlbs_dirty true and make_all_cpus_request.

All other memory barriers seems unnecessary. It can't be a invlpg relative to the page the mmu notifier is working on that is setting remote_tlbs_dirty after mmu_lock is released in the mmu notifier method, so as long as the remote_tlbs_dirty bit isn't lost we're ok. We basically have the remote_tlbs_dirty randomly going up from a mmu_lock protected section and the mmu notifier only must make sure that after clearing it, it always flushes (so only race is if we flush before clearing the flag).

However the setting must be done with set_bit and the clearing with test_and_clear_bit, otherwise when one cpus sets it inside mmu_lock protected section and other cpu clears it without any lock held from the mmu notifier, the result is undefined. AFIK without lock prefix on the mov instruction what can happen is like:

cpu0 cpu1
-------- ---------
remote_tlbs_dirty = 0
remote_tlbs_dirty = 1
remote_tlbs_dirty == 0

I don't think it only applies to bitops, surely it would more obviously apply to bitops but I think it also applies to plain mov.

I think we no longer need remote_tlbs_dirty. I'll send a new version.