Re: [patch 3/5] KVM: MMU: notifiers support for pinned sptes

[Gleb Natapov <gleb@kernel.org>]

On Wed, Jun 18, 2014 at 08:12:06PM -0300, mtosatti@redhat.com wrote:
> Request KVM_REQ_MMU_RELOAD when deleting sptes from MMU notifiers.
> 
> Keep pinned sptes intact if page aging.
> 
> Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
> 
> ---
>  arch/x86/kvm/mmu.c |   71 ++++++++++++++++++++++++++++++++++++++++++++++-------
>  1 file changed, 62 insertions(+), 9 deletions(-)
> 
> Index: kvm.pinned-sptes/arch/x86/kvm/mmu.c
> ===================================================================
> --- kvm.pinned-sptes.orig/arch/x86/kvm/mmu.c	2014-06-18 17:28:24.339435654 -0300
> +++ kvm.pinned-sptes/arch/x86/kvm/mmu.c	2014-06-18 17:29:32.510225755 -0300
> @@ -1184,6 +1184,42 @@
>  		kvm_flush_remote_tlbs(vcpu->kvm);
>  }
>  
> +static void ack_flush(void *_completed)
> +{
> +}
> +
> +static void mmu_reload_pinned_vcpus(struct kvm *kvm)
> +{
> +	int i, cpu, me;
> +	cpumask_var_t cpus;
> +	struct kvm_vcpu *vcpu;
> +	unsigned int req = KVM_REQ_MMU_RELOAD;
> +
> +	zalloc_cpumask_var(&cpus, GFP_ATOMIC);
> +
> +	me = get_cpu();
> +	kvm_for_each_vcpu(i, vcpu, kvm) {
> +		if (list_empty(&vcpu->arch.pinned_mmu_pages))
> +			continue;
> +		kvm_make_request(req, vcpu);
> +		cpu = vcpu->cpu;
> +
> +		/* Set ->requests bit before we read ->mode */
> +		smp_mb();
> +
> +		if (cpus != NULL && cpu != -1 && cpu != me &&
> +		      kvm_vcpu_exiting_guest_mode(vcpu) != OUTSIDE_GUEST_MODE)
> +			cpumask_set_cpu(cpu, cpus);
> +	}
> +	if (unlikely(cpus == NULL))
> +		smp_call_function_many(cpu_online_mask, ack_flush, NULL, 1);
> +	else if (!cpumask_empty(cpus))
> +		smp_call_function_many(cpus, ack_flush, NULL, 1);
> +	put_cpu();
> +	free_cpumask_var(cpus);
> +	return;
> +}
This is a c&p of make_all_cpus_request(), the only difference is checking
of vcpu->arch.pinned_mmu_pages.  You can add make_some_cpus_request(..., bool (*predicate)(struct kvm_vcpu *))
to kvm_main.c and rewrite make_all_cpus_request() to use it instead.

> +
>  /*
>   * Write-protect on the specified @sptep, @pt_protect indicates whether
>   * spte write-protection is caused by protecting shadow page table.
> @@ -1276,7 +1312,8 @@
>  }
>  
>  static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
> -			   struct kvm_memory_slot *slot, unsigned long data)
> +			   struct kvm_memory_slot *slot, unsigned long data,
> +			   bool age)
>  {
>  	u64 *sptep;
>  	struct rmap_iterator iter;
> @@ -1286,6 +1323,14 @@
>  		BUG_ON(!(*sptep & PT_PRESENT_MASK));
>  		rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", sptep, *sptep);
>  
> +		if (is_pinned_spte(*sptep)) {
> +			/* don't nuke pinned sptes if page aging: return
> + 			 * young=yes instead.
> + 			 */
> +			if (age)
> +				return 1;
> +			mmu_reload_pinned_vcpus(kvm);
> +		}
>  		drop_spte(kvm, sptep);
>  		need_tlb_flush = 1;
>  	}
> @@ -1294,7 +1339,8 @@
>  }
>  
>  static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
> -			     struct kvm_memory_slot *slot, unsigned long data)
> +			     struct kvm_memory_slot *slot, unsigned long data,
> +			     bool age)
>  {
>  	u64 *sptep;
>  	struct rmap_iterator iter;
> @@ -1312,6 +1358,9 @@
>  
>  		need_flush = 1;
>  
> +		if (is_pinned_spte(*sptep))
> +			mmu_reload_pinned_vcpus(kvm);
> +
>  		if (pte_write(*ptep)) {
>  			drop_spte(kvm, sptep);
>  			sptep = rmap_get_first(*rmapp, &iter);
> @@ -1342,7 +1391,8 @@
>  				int (*handler)(struct kvm *kvm,
>  					       unsigned long *rmapp,
>  					       struct kvm_memory_slot *slot,
> -					       unsigned long data))
> +					       unsigned long data,
> +					       bool age))
>  {
>  	int j;
>  	int ret = 0;
> @@ -1382,7 +1432,7 @@
>  			rmapp = __gfn_to_rmap(gfn_start, j, memslot);
>  
>  			for (; idx <= idx_end; ++idx)
> -				ret |= handler(kvm, rmapp++, memslot, data);
> +				ret |= handler(kvm, rmapp++, memslot, data, false);
>  		}
>  	}
>  
> @@ -1393,7 +1443,8 @@
>  			  unsigned long data,
>  			  int (*handler)(struct kvm *kvm, unsigned long *rmapp,
>  					 struct kvm_memory_slot *slot,
> -					 unsigned long data))
> +					 unsigned long data,
> +					 bool age))
>  {
>  	return kvm_handle_hva_range(kvm, hva, hva + 1, data, handler);
>  }
> @@ -1414,7 +1465,8 @@
>  }
>  
>  static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
> -			 struct kvm_memory_slot *slot, unsigned long data)
> +			 struct kvm_memory_slot *slot, unsigned long data,
> +			 bool age)
>  {
>  	u64 *sptep;
>  	struct rmap_iterator uninitialized_var(iter);
> @@ -1429,7 +1481,7 @@
>  	 * out actively used pages or breaking up actively used hugepages.
>  	 */
>  	if (!shadow_accessed_mask) {
> -		young = kvm_unmap_rmapp(kvm, rmapp, slot, data);
> +		young = kvm_unmap_rmapp(kvm, rmapp, slot, data, true);
>  		goto out;
>  	}
>  
> @@ -1450,7 +1502,8 @@
>  }
>  
>  static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
> -			      struct kvm_memory_slot *slot, unsigned long data)
> +			      struct kvm_memory_slot *slot, unsigned long data,
> +			      bool age)
>  {
>  	u64 *sptep;
>  	struct rmap_iterator iter;
> @@ -1488,7 +1541,7 @@
>  
>  	rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
>  
> -	kvm_unmap_rmapp(vcpu->kvm, rmapp, NULL, 0);
> +	kvm_unmap_rmapp(vcpu->kvm, rmapp, NULL, 0, false);
>  	kvm_flush_remote_tlbs(vcpu->kvm);
>  }
>  
> 
> 

--
			Gleb.