Re: [PATCH 6/6] Add age out of guest paused flag

[Eric B Munson]

Thanks for the review.

On Fri, 28 Oct 2011, Marcelo Tosatti wrote:

> On Tue, Oct 25, 2011 at 03:26:19PM -0400, Eric B Munson wrote:
> > The KVM_GUEST_PAUSED flag will prevent a guest from compaining about a soft
> > lockup but it can mask real soft lockups if the flag isn't cleared when it is
> > no longer relevant.  This patch adds a kvm ioctl that the hypervisor will use
> > when it resumes a guest to start a timer for aging out the flag.  The time out
> > will be specified by the hypervisor in the ioctl call.
> > 
> > Signed-off-by: Eric B Munson <emunson@xxxxxxxxx>
> > ---
> >  arch/x86/include/asm/pvclock.h |    2 ++
> >  arch/x86/kernel/kvmclock.c     |   24 ++++++++++++++++++++++++
> >  arch/x86/kvm/x86.c             |    9 +++++++++
> >  include/linux/kvm.h            |    2 ++
> >  include/linux/kvm_host.h       |    2 ++
> >  5 files changed, 39 insertions(+), 0 deletions(-)
> > 
> > diff --git a/arch/x86/include/asm/pvclock.h b/arch/x86/include/asm/pvclock.h
> > index 9312814..e8460b9 100644
> > --- a/arch/x86/include/asm/pvclock.h
> > +++ b/arch/x86/include/asm/pvclock.h
> > @@ -18,6 +18,8 @@ void kvm_set_host_stopped(struct kvm_vcpu *vcpu);
> >  
> >  bool kvm_check_and_clear_host_stopped(int cpu);
> >  
> > +void kvm_clear_guest_paused(struct kvm_vcpu *vcpu, unsigned int length);
> > +
> >  /*
> >   * Scale a 64-bit delta by scaling and multiplying by a 32-bit fraction,
> >   * yielding a 64-bit result.
> > diff --git a/arch/x86/kernel/kvmclock.c b/arch/x86/kernel/kvmclock.c
> > index f4fff3d..f3f3935 100644
> > --- a/arch/x86/kernel/kvmclock.c
> > +++ b/arch/x86/kernel/kvmclock.c
> > @@ -23,6 +23,8 @@
> >  #include <asm/apic.h>
> >  #include <linux/percpu.h>
> >  #include <linux/kvm_host.h>
> > +#include <linux/kobject.h>
> > +#include <linux/sysfs.h>
> >  
> >  #include <asm/x86_init.h>
> >  #include <asm/reboot.h>
> > @@ -144,6 +146,28 @@ bool kvm_check_and_clear_host_stopped(int cpu)
> >  	return ret;
> >  }
> >  
> > +static void kvm_timer_clear_guest_paused(unsigned long vcpu_addr)
> > +{
> > +	struct kvm_vcpu *vcpu = (struct kvm_vcpu *)vcpu_addr;
> > +	struct pvclock_vcpu_time_info *src = &vcpu->arch.hv_clock;
> > +	src->flags = src->flags & (!PVCLOCK_GUEST_STOPPED);
> > +}
> > +
> > +/*
> > + * Host has resumed the guest, we need to clear the guest paused flag so we
> > + * don't mask any real soft lockups.
> > + */
> > +void kvm_clear_guest_paused(struct kvm_vcpu *vcpu, unsigned int length)
> > +{
> > +	if (!timer_pending(&vcpu->flag_timer))
> > +		setup_timer(&vcpu->flag_timer,
> > +			    kvm_timer_clear_guest_paused,
> > +			    (unsigned long)vcpu);
> > +	mod_timer(&vcpu->flag_timer,
> > +		  jiffies + (length * HZ));
> > +}
> > +EXPORT_SYMBOL_GPL(kvm_clear_guest_paused);
> 
> So this is why the scheme could become awkward. You can't determine a   
> value for length that will guarantee that _no_ genuine softlockup is    
> omitted.
> 
> Can you think of a way to achieve this?

Given the way this set works now, I don't see a way to completely avoid the
chance of covering the first notification of a soft lockup.  But, if I
understand the way this code works, the watchdog will come back and complain
if the CPU is still soft locked on the next pass.  Unless I am not reading the
watchdog code correctly, this will only delay the warning of a real soft lockup.

> 
> Also note, all host code is located in arch/x86/kvm, guest 
> code in arch/x86/kernel/kvmclock.c.

I will reshuffle for the next revision.
>

Attachment: signature.asc
Description: Digital signature