Re: [PATCHV2 1/3] KVM: x86: introduce kvm_mmu_pte_prefetch structure

[Sergey Senozhatsky <senozhatsky@chromium.org>]

On (21/10/20 08:56), David Matlack wrote:
> > > The spinlock is overkill. I'd suggest something like this:
> > > - When VM-ioctl is invoked to update prefetch count, store it in
> > > kvm_arch. No synchronization with vCPUs needed.
> > > - When a vCPU takes a fault: Read the prefetch count from kvm_arch. If
> > > different than count at last fault, re-allocate vCPU prefetch array.
> > > (So you'll need to add prefetch array and count to kvm_vcpu_arch as
> > > well.)
> > >
> > > No extra locks are needed. vCPUs that fault after the VM-ioctl will
> > > get the new prefetch count. We don't really care if a prefetch count
> > > update races with a vCPU fault as long as vCPUs are careful to only
> > > read the count once (i.e. use READ_ONCE(vcpu->kvm.prefetch_count)) and
> > > use that. Assuming prefetch count ioctls are rare, the re-allocation
> > > on the fault path will be rare as well.
> >
> > So reallocation from the faul-path should happen before vCPU takes the
> > mmu_lock?
> 
> Yes. Take a look at mmu_topup_memory_caches for an example of
> allocating in the fault path prior to taking the mmu lock.
> 
> > And READ_ONCE(prefetch_count) should also happen before vCPU
> > takes mmu_lock, I assume, so we need to pass it as a parameter to all
> > the functions that will access prefetch array.
> 
> Store the value of READ_ONCE(prefetch_count) in struct kvm_vcpu_arch
> because you also need to know if it changes on the next fault. Then
> you also don't have to add a parameter to a bunch of functions in the
> fault path.
> 
> >
> > > Note: You could apply this same approach to a module param, except
> > > vCPUs would be reading the module param rather than vcpu->kvm during
> > > each fault.
> > >
> > > And the other alternative, like you suggested in the other patch, is
> > > to use a vCPU ioctl. That would side-step the synchronization issue
> > > because vCPU ioctls require the vCPU mutex. So the reallocation could
> > > be done in the ioctl and not at fault time.
> >
> > One more idea, wonder what do you think:
> >
> > There is an upper limit on the number of PTEs we prefault, which is 128 as of
> > now, but I think 64 will be good enough, or maybe even 32. So we can always
> > allocate MAX_PTE_PREFETCH_NUM arrays in vcpu->arch and ioctl() will change
> > ->num_ents only, which is always in (0, MAX_PTE_PREFETCH_NUM - 1] range. This
> > way we never have to reallocate anything, we just adjust the "maximum index"
> > value.
> 
> 128 * 8 would be 1KB per vCPU. That is probably reasonable, but I
> don't think the re-allocation would be that complex.

128 is probably too large. What I'm thinking is "32 * 8" per-VCPU.
Then it can even be something like:

---

diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 5271fce6cd65..b3a436f8fdf5 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -607,6 +607,11 @@ struct kvm_vcpu_xen {
        u64 runstate_times[4];
 };
 
+struct kvm_mmu_pte_prefetch {
+       u64 ents[32];
+       unsigned int num_ents;  /* max prefetch value for this vCPU */
+};
+
 struct kvm_vcpu_arch {
        /*
         * rip and regs accesses must go through
@@ -682,6 +687,8 @@ struct kvm_vcpu_arch {
        struct kvm_mmu_memory_cache mmu_gfn_array_cache;
        struct kvm_mmu_memory_cache mmu_page_header_cache;
 
+       struct kvm_mmu_pte_prefetch mmu_pte_prefetch;
+
        /*
         * QEMU userspace and the guest each have their own FPU state.
         * In vcpu_run, we switch between the user and guest FPU contexts.

---

> >
> > > Taking a step back, can you say a bit more about your usecase?
> >
> > We are looking at various ways of reducing the number of vm-exits. There
> > is only one VM running on the device (a pretty low-end laptop).
> 
> When you say reduce the number of vm-exits, can you be more specific?
> Are you trying to reduce the time it takes for vCPUs to fault in
> memory during VM startup?

VM Boot is the test I'm running, yes, and I see some improvements with
pte-prefault == 16.

> I just mention because there are likely other techniques you can apply
> that would not require modifying KVM code (e.g. prefaulting the host
> memory before running the VM, using the TDP MMU instead of the legacy
> MMU to allow parallel faults, using hugepages to map in more memory
> per fault, etc.)

THP would be awesome. We have THP enabled on devices that have 8-plus
gigabytes of RAM; but we don't have THP enabled on low end devices, that
only have 4 gigabytes of RAM. On low end devices KVM with THP causes host
memory regression, that we cannot accept, hence for 4 gig devices we try
various "other solutions".

We are using TDP. And somehow I never see (literally never) async PFs.
It's always either hva_to_pfn_fast() or hva_to_pfn_slow() or
__direct_map() from tdp_page_fault().