Re: [PATCH 10/21] KVM: x86/mmu: Make __GFP_ZERO a property of the memory cache

From: Sean Christopherson
Date: Mon Jun 22 2020 - 15:41:00 EST


On Wed, Jun 10, 2020 at 11:57:32AM -0700, Ben Gardon wrote:
> > ---
> > arch/x86/include/asm/kvm_host.h | 1 +
> > arch/x86/kvm/mmu/mmu.c | 7 ++++++-
> > 2 files changed, 7 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> > index e7a427547557..fb99e6776e27 100644
> > --- a/arch/x86/include/asm/kvm_host.h
> > +++ b/arch/x86/include/asm/kvm_host.h
> > @@ -251,6 +251,7 @@ struct kvm_kernel_irq_routing_entry;
> > */
> > struct kvm_mmu_memory_cache {
> > int nobjs;
> > + gfp_t gfp_zero;
> This would make more sense to me if it could be used for general extra
> gfp flags and was called gfp_flags or something, or it was a boolean
> that was later translated into the flag being set. Storing the
> gfp_zero flag here is a little counter-intuitive. Probably not worth
> changing unless you're sending out a v2 for some other reason.

Ideally, this would be a generic gfp_flags field, but that's basically a
non-starter for patch 5, which uses GFP_ATOMIC for the "oh crap the cache
is empty" error handling. Allowing arbitrary flags would be a mess.

I went with storing a full gfp_t because that produces more optimal code.
This isn't a super critical path and it's only a few cycles, but it seems
worthwhile given the frequency with which this code will be called, and
since this happens under mmu_lock.


348 gfp_flags |= mc->gfp_zero;
0x00000000000058ab <+59>: mov 0x4(%rbx),%eax
0x00000000000058ae <+62>: or $0x400cc0,%eax

versus

349 gfp_flags |= __GFP_ZERO;
0x00000000000058a7 <+55>: cmpb $0x1,0x4(%rbx)
0x00000000000058ab <+59>: mov 0x8(%rbx),%rdi <-- unrelated interleaved code
0x00000000000058af <+63>: sbb %eax,%eax
0x00000000000058b1 <+65>: xor %al,%al
0x00000000000058b3 <+67>: add $0x400dc0,%eax