Re: [PATCH RFC 01/15] KVM: Move running VCPU from ARM to common code
From: Peter Xu
Date: Mon Dec 09 2019 - 17:05:19 EST
On Wed, Dec 04, 2019 at 10:42:27AM +0100, Paolo Bonzini wrote:
> On 03/12/19 20:01, Sean Christopherson wrote:
> > In case it was clear, I strongly dislike adding kvm_get_running_vcpu().
> > IMO, it's a unnecessary hack. The proper change to ensure a valid vCPU is
> > seen by mark_page_dirty_in_ring() when there is a current vCPU is to
> > plumb the vCPU down through the various call stacks. Looking up the call
> > stacks for mark_page_dirty() and mark_page_dirty_in_slot(), they all
> > originate with a vcpu->kvm within a few functions, except for the rare
> > case where the write is coming from a non-vcpu ioctl(), in which case
> > there is no current vCPU.
> > The proper change is obviously much bigger in scope and would require
> > touching gobs of arch specific code, but IMO the end result would be worth
> > the effort. E.g. there's a decent chance it would reduce the API between
> > common KVM and arch specific code by eliminating the exports of variants
> > that take "struct kvm *" instead of "struct kvm_vcpu *".
> It's not that simple. In some cases, the "struct kvm *" cannot be
> easily replaced with a "struct kvm_vcpu *" without making the API less
> intuitive; for example think of a function that takes a kvm_vcpu pointer
> but then calls gfn_to_hva(vcpu->kvm) instead of the expected
> That said, looking at the code again after a couple years I agree that
> the usage of kvm_get_running_vcpu() is ugly. But I don't think it's
> kvm_get_running_vcpu()'s fault, rather it's the vCPU argument in
> mark_page_dirty_in_slot and mark_page_dirty_in_ring that is confusing
> and we should not be adding.
> kvm_get_running_vcpu() basically means "you can use the per-vCPU ring
> and avoid locking", nothing more. Right now we need the vCPU argument
> in mark_page_dirty_in_ring for kvm_arch_vcpu_memslots_id(vcpu), but that
> is unnecessary and is the real source of confusion (possibly bugs too)
> if it gets out of sync.
> Instead, let's add an as_id field to struct kvm_memory_slot (which is
> trivial to initialize in __kvm_set_memory_region), and just do
> as_id = slot->as_id;
> vcpu = kvm_get_running_vcpu();
> in mark_page_dirty_in_ring.
Looks good. I'm adding another patch for it, and dropping patch 2 then.