[Devel] [PATCH rh7] ve: Add interface for @start_timespec and @real_start_timespec adjustmen

[Vladimir Davydov]

On Tue, Feb 09, 2016 at 01:53:04PM +0300, Cyrill Gorcunov wrote:
> On Tue, Feb 09, 2016 at 01:29:56PM +0300, Vladimir Davydov wrote:
> > > 
> > > Yes. We take current values then substract the @real_[start_]timespec
> > > and return it. IOW they are just monotonic|boottime clocks.
> > 
> > Then I think we should name these files accordingly to avoid confusion.
> > ve.clock_{monotonic,boottime} probably.
> 
> It was named so in pcs6, so I tried to preserve naming. But sure I
> can rename, that's not a problem.

I mean not ve_struct.start_timespec - its name is perfectly fine - but
the cgroup file ve.start_timespec - with your patch it should rather be
called ve.clock_monotonic.

> 
> > > > This looks weird to me. Wouldn't it be better if we returned the
> > > > timespec as is?
> > > 
> > > No, we have to account real total_sleep_time the container
> > > has on the node, because when you close the lid the total
> > > sleep time will increse but @real_start_timespec won't
> > > as far as I can say and this gonna be wrong.
> > 
> > But you can always calculate the difference in userspace, like
> > (CLOCK_BOOTTIME - ve.real_start_timespec), no?
> 
> No I fear not. Look, the calculation of how long container has been
> uptime is fully up to the kernel, the @real_[start_]timespec is
> not update by the kernel during clock ticks (this is completely
> done in timekeeper code).
> 
> Vladimir, maybe I misunderstand you, could you please provide
> the pseudocode how you imagine it could work? I mean without
> in-kernel time normalization. I suspect I miss something obvious.

Before suspending a container, you call clock_gettime for
CLOCK_MONOTONIC on the host, then read ve.start_timespec, then save the
difference to the image. On restore you read CLOCK_MONOTONIC on the
destination host and write (clock - difference) to ve.start_timepsec.
This would require more work from userspace though.