On Fri, Jun 5, 2020 at 2:39 PM Sean Christopherson
<sean.j.christopherson@xxxxxxxxx> wrote:
>
> Attempt to allocate a new object instead of crashing KVM (and likely the
> kernel) if a memory cache is unexpectedly empty. Use GFP_ATOMIC for the
> allocation as the caches are used while holding mmu_lock. The immediate
> BUG_ON() makes the code unnecessarily explosive and led to confusing
> minimums being used in the past, e.g. allocating 4 objects where 1 would
> suffice.
>
> Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> ---
> arch/x86/kvm/mmu/mmu.c | 21 +++++++++++++++------
> 1 file changed, 15 insertions(+), 6 deletions(-)
>
> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> index ba70de24a5b0..5e773564ab20 100644
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -1060,6 +1060,15 @@ static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
> local_irq_enable();
> }
>
> +static inline void *mmu_memory_cache_alloc_obj(struct kvm_mmu_memory_cache *mc,
> + gfp_t gfp_flags)
> +{
> + if (mc->kmem_cache)
> + return kmem_cache_zalloc(mc->kmem_cache, gfp_flags);
> + else
> + return (void *)__get_free_page(gfp_flags);
> +}
> +
> static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *mc, int min)
> {
> void *obj;
> @@ -1067,10 +1076,7 @@ static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *mc, int min)
> if (mc->nobjs >= min)
> return 0;
> while (mc->nobjs < ARRAY_SIZE(mc->objects)) {
> - if (mc->kmem_cache)
> - obj = kmem_cache_zalloc(mc->kmem_cache, GFP_KERNEL_ACCOUNT);
> - else
> - obj = (void *)__get_free_page(GFP_KERNEL_ACCOUNT);
> + obj = mmu_memory_cache_alloc_obj(mc, GFP_KERNEL_ACCOUNT);
> if (!obj)
> return mc->nobjs >= min ? 0 : -ENOMEM;
> mc->objects[mc->nobjs++] = obj;
> @@ -1118,8 +1124,11 @@ static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
> {
> void *p;
>
> - BUG_ON(!mc->nobjs);
> - p = mc->objects[--mc->nobjs];
> + if (WARN_ON(!mc->nobjs))
> + p = mmu_memory_cache_alloc_obj(mc, GFP_ATOMIC | __GFP_ACCOUNT);
Is an atomic allocation really necessary here? In most cases, when
topping up the memory cache we are handing a guest page fault. This
bug could also be removed by returning null if unable to allocate from
the cache, and then re-trying the page fault in that case. I don't
know if this is necessary to handle other, non-x86 architectures more
easily, but I worry this could cause some unpleasantness if combined
with some other bug or the host was in a low memory situation and then
this consumed the atomic pool. Perhaps this is a moot point since we
log a warning and consider the atomic allocation something of an
error.
> + else
> + p = mc->objects[--mc->nobjs];
> + BUG_ON(!p);
> return p;
> }
>
> --
> 2.26.0
>
