Re: [PATCH v2 1/5] KVM: x86: Introduce KVM_REQ_RING_SOFT_FULL

[Marc Zyngier]

On Mon, 19 Sep 2022 00:58:10 +0100,
Gavin Shan <gshan@xxxxxxxxxx> wrote:
> 
> On 9/18/22 7:00 PM, Marc Zyngier wrote:
> > On Fri, 16 Sep 2022 19:09:52 +0100,
> > Peter Xu <peterx@xxxxxxxxxx> wrote:
> >> 
> >> On Fri, Sep 16, 2022 at 12:51:31PM +0800, Gavin Shan wrote:
> >>> This adds KVM_REQ_RING_SOFT_FULL, which is raised when the dirty
> >>> ring of the specific VCPU becomes softly full in kvm_dirty_ring_push().
> >>> The VCPU is enforced to exit when the request is raised and its
> >>> dirty ring is softly full on its entrance.
> >>> 
> >>> Suggested-by: Marc Zyngier <maz@xxxxxxxxxx>
> >>> Signed-off-by: Gavin Shan <gshan@xxxxxxxxxx>
> >>> ---
> >>>   arch/x86/kvm/x86.c       | 5 +++--
> >>>   include/linux/kvm_host.h | 1 +
> >>>   virt/kvm/dirty_ring.c    | 4 ++++
> >>>   3 files changed, 8 insertions(+), 2 deletions(-)
> >>> 
> >>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> >>> index 43a6a7efc6ec..7f368f59f033 100644
> >>> --- a/arch/x86/kvm/x86.c
> >>> +++ b/arch/x86/kvm/x86.c
> >>> @@ -10265,8 +10265,9 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
> >>>   	bool req_immediate_exit = false;
> >>>     	/* Forbid vmenter if vcpu dirty ring is soft-full */
> >>> -	if (unlikely(vcpu->kvm->dirty_ring_size &&
> >>> -		     kvm_dirty_ring_soft_full(&vcpu->dirty_ring))) {
> >>> +	if (kvm_check_request(KVM_REQ_RING_SOFT_FULL, vcpu) &&
> >>> +	    kvm_dirty_ring_soft_full(&vcpu->dirty_ring)) {
> >>> +		kvm_make_request(KVM_REQ_RING_SOFT_FULL, vcpu);
> >>>   		vcpu->run->exit_reason = KVM_EXIT_DIRTY_RING_FULL;
> >>>   		trace_kvm_dirty_ring_exit(vcpu);
> >>>   		r = 0;
> >> 
> >> As commented previously - can we use kvm_test_request() instead? because we
> >> don't want to unconditionally clear the bit.  Instead of making the request
> >> again, we can clear request only if !full.
> > 
> > I have the feeling that this is a micro-optimisation that won't lead
> > to much benefit in practice. You already have the cache line, just not
> > in exclusive mode, and given that this is per-vcpu, you'd only see the
> > cost if someone else is setting a request to this vcpu while
> > evaluating the local requests.
> > 
> > And now you need extra barriers...
> > 
> > Also, can we please refrain from changing things without data showing
> > that this actually is worse than what we had before? The point below
> > makes me think that this is actually beneficial as is.
> > 
> 
> I think Marc's explanation makes sense. It won't make difference in terms
> of performance. We need to explicitly handle barrier when kvm_test_request()
> is used. So I prefer to keep the code if Peter agrees.
> 
> >> We can also safely move this into the block of below kvm_request_pending()
> >> as Marc used to suggest.
> > 
> > This, on the other hand, makes sure that we share the cost across all
> > requests. Requests should be extremely rare anyway (and if they
> > aren't, you have a whole lot of performance issues on your hands
> > anyway).
> > 
> 
> Yeah, We shouldn't have too much requests. I missed the comment from Marc
> to move this chunk to kvm_request_pending(). I will fix it in v3.
> 
> >> 
> >> To explicitly use kvm_clear_request(), we may need to be careful on the
> >> memory barriers.  I'm wondering whether we should have moved
> >> smp_mb__after_atomic() into kvm_clear_request() because kvm_clear_request()
> >> is used outside kvm_check_request() and IIUC all the call sites should
> >> better have that barrier too to be safe.
> >> 
> >> Side note: when I read the code around I also see some mis-use of clear
> >> request where it can be omitted, e.g.:
> >> 
> >> 		if (kvm_check_request(KVM_REQ_UNHALT, vcpu)) {
> >> 			kvm_clear_request(KVM_REQ_UNHALT, vcpu);
> >> 			vcpu->run->exit_reason = KVM_EXIT_IRQ_WINDOW_OPEN;
> >> 		}
> >> 
> >> Maybe it's a sign of bad naming, so we should renamed kvm_check_request()
> >> to kvm_test_clear_request() too to show that clearing after that is not
> >> needed?
> > 
> > Yeah, this kvm_clear_request() is superfluous. But this is rather well
> > documented, for once, and I don't think we should repaint it based on
> > a sample of one.
> > 
> 
> Yeah, I think Peter is correct that smp_mb__after_atomic() would be
> part of kvm_clear_request(). Otherwise, the following two cases aren't
> in same order:
> 
>       // kvm_check_request()             // test and clear
>       kvm_test_request()                 kvm_test_request()
>       kvm_clear_request()                kvm_clear_request()
>       smp_mb__after_atomic()

[fixing Drew's email, again]

Oh, I totally agree that a standalone use of kvm_clear_request() must
come with a barrier. It is just that without additional data, it isn't
obvious to me that there is any need for kvm_clear_request() itself to
have the barrier.

In a number of cases, kvm_clear_request() is used on in the context of
the same vcpu, and this should be enough to ensure visibility (for
example, I don't think kvm_vcpu_wfi() should require this barrier).

But maybe I'm missing something.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.