Re: [RFC/PATCH] SLQB: Mark the allocator as broken PowerPC and S390

[Mel Gorman <mel@csn.ul.ie>]

On Thu, Sep 17, 2009 at 02:13:39PM +0300, Pekka Enberg wrote:
> Hi Mel,
> 
> On Thu, 2009-09-17 at 11:57 +0100, Mel Gorman wrote:
> > The danger is that this isn't a PPC or s390 bug then as such, but a bug where
> > there are either memoryless nodes or when node 0 is memoryless.  Hence, there
> > is no guarantee that your Kconfig option will catch all instances where this
> > bug triggers.  Granted, the configuration is most likely a PPC machine :)
> 
> Yes, I suggested making SLQB depend on !NUMA to Nick but he didn't like
> that as it's known to be good on x86 NUMA configs.
> 

Agreed.

> On Thu, 2009-09-17 at 11:08 +0100, Mel Gorman wrote:
> > > > The danger is if SLQB is being silently disabled, it'll never be noticed
> > > > or debugged :/
> > > 
> > > Maybe, but that's not an excuse to push something that's known to break. 
> 
> On Thu, 2009-09-17 at 11:57 +0100, Mel Gorman wrote:
> > Wow, this is from back in May! Lame.
> 
> Heh, my (lame) excuse is lack of relevant hardware.... ;-)
> 

I'm not blaming you. It's just ... unfortunate :/

> On Thu, 2009-09-17 at 11:57 +0100, Mel Gorman wrote:
> > I'm against silently disabling it. Memoryless nodes are extremely rare but
> > bugs crop up there occasionally and take a long time to catch and squash. SLQB
> > breaking there is not going to cause widespread damage but force a fix to
> > be developed by the people with access to the affected machines.
> 
> Hey, if someone sends me fix for the bug well before the merge window
> closes, that would be great! But there's no way we're adding new core
> kernel code that's _known_ to break peoples configs, at least not
> through slab.git. If disabling SLQB is not acceptable and we're unable
> to fix things, we'll just have to skip this release cycle.
> 

Please consider disabling it as an option in the rc3 stage or the like.
With luck, I'll find a suitable machine in time and see what can be
done. I just don't like the idea of x86 defaulting to one allocator and
ppc defaulting to another.

> On Thu, 2009-09-17 at 11:57 +0100, Mel Gorman wrote:
> > Total aside, does anybody know handily if fake NUMA support allows the
> > creation of memoryless nodes help reproducing problems like this? If I can't
> > get a real machine, that'll be the approach I'll be trying.
> 
> That would be useful, yes.
> 

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab