Re: v4.10-rc8 (-rc6) boot regression on Intel desktop, does not boot after cold boots, boots after reboot

[Thomas Gleixner]

On Fri, 17 Feb 2017, Frederic Weisbecker wrote:
> On Thu, Feb 16, 2017 at 08:34:45PM +0100, Thomas Gleixner wrote:
> > On Thu, 16 Feb 2017, Frederic Weisbecker wrote:
> > > On Thu, Feb 16, 2017 at 10:20:14AM -0800, Linus Torvalds wrote:
> > > > On Thu, Feb 16, 2017 at 10:13 AM, Frederic Weisbecker
> > > > <fweisbec@xxxxxxxxx> wrote:
> > > > >
> > > > > I haven't followed the discussion but this patch has a known issue which is fixed
> > > > > with:
> > > > >     7bdb59f1ad474bd7161adc8f923cdef10f2638d1
> > > > >     "tick/nohz: Fix possible missing clock reprog after tick soft restart"
> > > > >
> > > > > I hope this fixes your issue.
> > > > 
> > > > No, Pavel saw the problem with rc8 too, which already has that fix.
> > > > 
> > > > So I think we'll just need to revert that original patch (and that
> > > > means that we have to revert the commit you point to as well, since
> > > > that ->next_tick field was added by the original commit).
> > > 
> > > Aw too bad, but indeed that late we don't have the choice.
> > 
> > Hint: Look for CPU hotplug interaction of these patches. I bet something
> > becomes stale when the CPU goes down and does not get reset when it comes
> > back online.
> 
> Indeed I should check that. But Pavel is seeing this on boot, where the

I don't think so. He observed it on suspend resume and by doing hotplug
operations in a loop. But I might be wrong as usual.

> only hotplug operations that happen are CPU UP without preceding CPU DOWN
> that may have retained stale values. I think the value of ts->next_tick should
> be initially 0 for all CPUs. So perhaps that 0 value confuses stuff. But
> looking at the code I don't see how. It maybe something more subtle.
>