Re: KVM: x86: update masterclock when kvmclock_offset is calculated

[Marcelo Tosatti <mtosatti@redhat.com>]

On Tue, Sep 03, 2013 at 12:42:40PM +0200, Paolo Bonzini wrote:
> Il 03/09/2013 05:03, Marcelo Tosatti ha scritto:
> > > Here is the code I was thinking of:
> > > 
> > > 	spin_lock(&ka->pvclock_gtod_sync_lock);
> > > 	make_all_cpus_request(kvm, KVM_REQ_CLOCK_UPDATE);
> > > 
> > > 	/*
> > > 	 * No guest entries from this point: VCPUs will be spinning
> > > 	 * on pvclock_gtod_sync_lock in kvm_guest_time_update.
> > > 	 */
> > > 	pvclock_update_vm_gtod_copy(kvm);
> > > 
> > > 	/*
> > > 	 * Let kvm_guest_time_update continue: entering the guest
> > > 	 * is now allowed too.
> > > 	 */
> > > 	spin_unlock(&ka->pvclock_gtod_sync_lock);
> > > 
> > > KVM_REQ_CLOCK_UPDATE is used to cancel guest entry and execute
> > > kvm_guest_time_update.  But kvm_guest_time_update will spin on
> > > pvclock_gtod_sync_lock until pvclock_update_vm_gtod_copy exits and
> > > kvm_gen_update_masterclock releases the spinlock.
> > 
> > Not safe because there are places which set KVM_REQ_CLOCK_UPDATE without
> > kicking target vcpu out of guest mode. Unless you use a modified
> > make_all_cpus_request.
> 
> make_all_cpus_request does force an exit, even if the bit is already set
> in vcpus->requests.  But maybe here is where I'm missing something.

No, you're right. My bad.

> > The point of REQ_MCLOCK_INPROGRESS request is to guarantee that the
> > following is not possible:
> > 
> > - 2 vcpus in guest mode with per-vcpu kvmclock areas with
> > different {system_timestamp, tsc_offset} values.
> 
> Understood.
> 
> > To achieve that:
> > 
> > - Kick all vcpus out of guest mode (via a request bit that can't be
> >   cleared).
> > - Update the {system_timestamp, tsc_offset} values.
> > - Clear the request bit.
> 
> After make_all_cpus_request, all VCPUs will be out of guest mode, and
> the request bit will not be cleared until pvclock_gtod_sync_lock is
> released.
> 
> Paolo

It seems more obfuscated to rely on an implicit pvclock_gtod_sync_lock
blocking than an explicit request bit that can't be cleared, but perhaps
thats personal opinion.

OTOH, you can also argue that the request bit abuses the vcpu->requests
request mechanism, because there is not a request in fact (it only
blocks guest entry).

If you have reasons to update, and you are sure its safe, feel
free to modify it.