Re: [PATCH] kvm/debugfs: add file to get vcpu steal time statistics

From: Sean Christopherson
Date: Mon Sep 30 2024 - 11:45:44 EST


On Mon, Sep 30, 2024, Denis Plotnikov wrote:
>
>
> On 9/23/24 14:46, Sean Christopherson wrote:
> > On Mon, Sep 23, 2024, Denis Plotnikov wrote:
> > > On 9/22/24 11:04, Sean Christopherson wrote:
> > > > On Tue, Sep 17, 2024, Denis Plotnikov wrote:
> > > > > It's helpful to know whether some other host activity affects a virtual
> > > > > machine to estimate virtual machine quality of sevice.
> > > > > The fact of virtual machine affection from the host side can be obtained
> > > > > by reading "preemption_reported" counter via kvm entries of sysfs, but
> > > > > the exact vcpu waiting time isn't reported to the host.
> > > > > This patch adds this reporting.
> > > > >
> > > > > Signed-off-by: Denis Plotnikov <den-plotnikov@xxxxxxxxxxxxxx>
> > > > > ---
> > > > > arch/x86/include/asm/kvm_host.h | 1 +
> > > > > arch/x86/kvm/debugfs.c | 17 +++++++++++++++++
> > > >
> > > > Using debugfs is undesirable, as it's (a) not ABI and (b) not guaranteed to be
> > > > present as KVM (correctly) ignores debugfs setup errors.
> > > >
> > > > Using debugfs is also unnecessary. The total steal time is available in guest
> > > > memory, and by definition that memory is shared with the host. To query total
> > > > steal time from userspace, use MSR filtering to trap writes (and reflect writes
> > > > back into KVM) so that the GPA of the steal time structure is known, and then
> > > > simply read the actual steal time from guest memory as needed.
> > > Thanks for the reply!
> > > Just to clarify, by reading the actual steal time from guest memory do you
> > > mean by using some kind of new vcpu ioctl?
> >
> > No, I mean by using the host userspace VMA to read the memory.
>
> Oh, I think I got your idea. You mean
> using KVM_CAP_X86_MSR_FILTER which...
>
> "In combination with KVM_CAP_X86_USER_SPACE_MSR, this allows user space to
> trap and emulate MSRs ..."
>
> And then having guest's steal time struct valid address read the value from
> userspace VMM like qemu directly.

Yep, exactly!