Re: [RFC PATCH 6/9] livepatch: create per-task consistency model

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Thu, Feb 12, 2015 at 01:42:01PM +0100, Jiri Kosina wrote:
> On Thu, 12 Feb 2015, Peter Zijlstra wrote:
> 
> > > Well, the fact indisputable fact is that there is a demand for this. It's 
> > > not about one machine, it's about scheduling dowtimes of datacentres.
> > 
> > The changelog says:
> > 
> >  > ... A patch can remain in the
> >  > transition state indefinitely, if any of the tasks are stuck in the
> >  > previous universe.
> > 
> > Therefore there is no scheduling anything. Without timeliness guarantees
> > you can't make a schedule.
> > 
> > Might as well just reboot, at least that's fairly well guaranteed to
> > happen.
> 
> All running (reasonably alive) tasks will be running patched code though. 
> 
> You can't just claim complete victory (and get ready for accepting another 
> patch, etc) if there is a long-time sleeper that hasn't been converted 
> yet.

Agreed.  And also we have several strategies for reducing the time
needed to get all tasks to a patched state (see patch 9 of this series
for more details).  The goal is to not leave systems in limbo for more
than a few seconds.

-- 
Josh