On Thu, Jun 20, 2013 at 09:28:37PM +0900, Takuya Yoshikawa wrote:
> On Thu, 20 Jun 2013 14:45:04 +0300
> Gleb Natapov <g...@redhat.com> wrote:
>
> > On Thu, Jun 20, 2013 at 12:59:54PM +0200, Paolo Bonzini wrote:
> > > Il 20/06/2013 10:59, Takuya Yoshikawa ha scritto:
> > > > Without this information, users will just see unexpected performance
> > > > problems and there is little chance we will get good reports from them:
> > > > note that mmio generation is increased even when we just start, or stop,
> > > > dirty logging for some memory slot, in which case users should never
> > > > expect all shadow pages to be zapped.
> > > >
> > > > Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya...@lab.ntt.co.jp>
> > > > ---
> > > > arch/x86/kvm/mmu.c | 4 +++-
> > > > 1 file changed, 3 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> > > > index c60c5da..bc8302f 100644
> > > > --- a/arch/x86/kvm/mmu.c
> > > > +++ b/arch/x86/kvm/mmu.c
> > > > @@ -4385,8 +4385,10 @@ void kvm_mmu_invalidate_mmio_sptes(struct kvm
> > > > *kvm)
> > > > * The max value is MMIO_MAX_GEN - 1 since it is not called
> > > > * when mark memslot invalid.
> > > > */
> > > > - if (unlikely(kvm_current_mmio_generation(kvm) >= (MMIO_MAX_GEN
> > > > - 1)))
> > > > + if (unlikely(kvm_current_mmio_generation(kvm) >= (MMIO_MAX_GEN
> > > > - 1))) {
> > > > + printk(KERN_INFO "kvm: zapping shadow pages for mmio
> > > > generation wraparound");
> > >
> > > This should at least be rate-limited, because it is guest triggerable.
> > >
> > It will be hard for guest to triggers it 1 << 19 times too fast though.
>
> I think guest-triggerable zap_all itself is a threat for the host, rather
> than a matter of log flooding, even if it can be preempted.
>
It's not much we can do about it. Slot removal/creation is triggerable
through HW emulation registers.
> >
> > > But why isn't the kvm_mmu_invalidate_zap_all_pages tracepoint enough?
> > >
> > This one will trigger during slot deletion/move too.
> >
> > I would put it in to see if it actually triggers in some real world
> > workloads (skipping the firs wraparound since it is intentional),
> > we can always drop it if it will turn out to create a lot of noise.
> >
>
> This patch is not for developers but for end users: of course they do not
> use tracers during running their services normally.
>
> If they see mysterious peformance problems induced by this wraparound, the
> only
> way to know the cause later is by this kind of information in the syslog.
> So even the first wraparound may better be printed out IMO.
Think about starting hundreds VMs on a freshly booted host. You will see
hundreds of those pretty quickly.
>
> I want to let administrators know the cause if possible, any better way?
>
Not that I can think of. Paolo what about print_once() and ignore first
wraparound?
--
Gleb.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
