Re: [PATCH 2/2] mm: page allocator: Do not drain per-cpu lists via IPI from page allocator context

[Mel Gorman <mgorman@suse.de>]

On Fri, Jan 20, 2012 at 03:16:58AM +0530, Srivatsa S. Bhat wrote:
> [Reinstating the original Cc list]
> 
> On 01/19/2012 09:50 PM, Mel Gorman wrote:> 
> 
> > On a different x86-64 machines with an intel-specific MCE, I have
> > also noted that the value of num_online_cpus() can change while
> > stop_machine() is running.
> 
> 
> That is expected and intentional right? Meaning, it is during the
> stop_machine() thing itself that a CPU is actually taken offline.
> And at the same time, it is removed from the cpu_online_mask.
> 

It's intentional sometimes and no others. The machine does halt
sometimes and stays there.

> On Intel boxes, essentially, the following gets executed on the dying
> CPU, as set up by the stop_machine stuff.
> 
> __cpu_disable()
>     native_cpu_disable()
>         cpu_disable_common()
>             remove_cpu_from_maps()
>                 set_cpu_online(cpu, false)
> 			^^^^^^
> So, set_cpu_online will remove this CPU from the cpu_online_mask.
> And all this runs while still under the stop machine context.
> And this is exactly what we want right?
> 

We don't want it to halt in stop_machine forever waiting on acknowledges
that are never received until the NMI handler fires.

> > This is sensitive to timing and part of
> > the problem seems to be due to cmci_rediscover() running without the
> > CPU hotplug mutex held. This is not related to the IPI mess and is
> > unrelated to memory pressure but is just to note that CPU hotplug in
> > general can be fragile in parts.
> > 
> 
> 
> For the cmci_rediscover() part, I feel a simple get/put_online_cpus()
> around it should work.
> 

Yeah, that's the first thing I tried first too. Doesn't work though.

-- 
Mel Gorman
SUSE Labs