On Fri, 20 Mar 2020, Tom Lendacky wrote:
> Currently, CLFLUSH is used to flush SEV guest memory before the guest is
> terminated (or a memory hotplug region is removed). However, CLFLUSH is
> not enough to ensure that SEV guest tagged data is flushed from the cache.
>
> With 33af3a7ef9e6 ("KVM: SVM: Reduce WBINVD/DF_FLUSH invocations"), the
> original WBINVD was removed. This then exposed crashes at random times
> because of a cache flush race with a page that had both a hypervisor and
> a guest tag in the cache.
>
> Restore the WBINVD when destroying an SEV guest and add a WBINVD to the
> svm_unregister_enc_region() function to ensure hotplug memory is flushed
> when removed. The DF_FLUSH can still be avoided at this point.
>
> Fixes: 33af3a7ef9e6 ("KVM: SVM: Reduce WBINVD/DF_FLUSH invocations")
> Signed-off-by: Tom Lendacky <thomas.lendacky@xxxxxxx>
Acked-by: David Rientjes <rientjes@xxxxxxxxxx>
Should this be marked for stable?
Cc: stable@xxxxxxxxxxxxxxx # 5.5+
> ---
> arch/x86/kvm/svm.c | 22 ++++++++++++++--------
> 1 file changed, 14 insertions(+), 8 deletions(-)
>
> diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
> index 08568ae9f7a1..d54cdca9c140 100644
> --- a/arch/x86/kvm/svm.c
> +++ b/arch/x86/kvm/svm.c
> @@ -1980,14 +1980,6 @@ static void sev_clflush_pages(struct page *pages[], unsigned long npages)
> static void __unregister_enc_region_locked(struct kvm *kvm,
> struct enc_region *region)
> {
> - /*
> - * The guest may change the memory encryption attribute from C=0 -> C=1
> - * or vice versa for this memory range. Lets make sure caches are
> - * flushed to ensure that guest data gets written into memory with
> - * correct C-bit.
> - */
> - sev_clflush_pages(region->pages, region->npages);
> -
> sev_unpin_memory(kvm, region->pages, region->npages);
> list_del(®ion->list);
> kfree(region);
> @@ -2004,6 +1996,13 @@ static void sev_vm_destroy(struct kvm *kvm)
>
> mutex_lock(&kvm->lock);
>
> + /*
> + * Ensure that all guest tagged cache entries are flushed before
> + * releasing the pages back to the system for use. CLFLUSH will
> + * not do this, so issue a WBINVD.
> + */
> + wbinvd_on_all_cpus();
> +
> /*
> * if userspace was terminated before unregistering the memory regions
> * then lets unpin all the registered memory.
> @@ -7247,6 +7246,13 @@ static int svm_unregister_enc_region(struct kvm *kvm,
> goto failed;
> }
>
> + /*
> + * Ensure that all guest tagged cache entries are flushed before
> + * releasing the pages back to the system for use. CLFLUSH will
> + * not do this, so issue a WBINVD.
> + */
> + wbinvd_on_all_cpus();
> +
> __unregister_enc_region_locked(kvm, region);
>
> mutex_unlock(&kvm->lock);
> --
> 2.17.1
>
>
