Re: [BUG] THP allocations escape cpuset when defrag is off

[Alex Thorlton]

On Wed, Jul 23, 2014 at 03:28:09PM -0700, David Rientjes wrote:
> > My debug code shows that certain code paths are still allowing
> > ALLOC_CPUSET to get pulled off the alloc_flags with the patch, but
> > monitoring the memory usage shows that we're staying on node, aside from
> > some very small allocations, which may be other types of allocations that
> > are not necessarly confined to a cpuset.  Need a bit more research to
> > confirm that.
> > 
> 
> ALLOC_CPUSET should get stripped for the cases outlined in 
> __cpuset_node_allowed_softwall(), specifically for GFP_ATOMIC which does 
> not have __GFP_WAIT set.

Makes sense.  I knew my patch was probably the wrong way to fix this,
but it did serve my purpose :)

> > So, my question ends up being, why do we wipe out ___GFP_WAIT when
> > defrag is off?  I'll trust that there is good reason to do that, but, if
> > so, is the behavior that I'm seeing expected?
> > 
> 
> The intention is to avoid memory compaction (and direct reclaim), 
> obviously, which does not run when __GFP_WAIT is not set.  But you're 
> exactly right that this abuses the allocflags conversion that allows 
> ALLOC_CPUSET to get cleared because it is using the aforementioned 
> GFP_ATOMIC exception for cpuset allocation.
>
> We can't use PF_MEMALLOC or TIF_MEMDIE for hugepage allocation because it 
> affects the allowed watermarks and nothing else prevents memory compaction 
> or direct reclaim from running in the page allocator slowpath.
> 
> So it looks like a modification to the page allocator is needed, see 
> below.

Looks good to me.  Fixes the problem without affecting any of the other
intended functionality.

> It's also been a long-standing issue that cpusets and mempolicies are 
> ignored by khugepaged that allows memory to be migrated remotely to nodes 
> that are not allowed by a cpuset's mems or a mempolicy's nodemask.  Even 
> with this issue fixed, you may find that some memory is migrated remotely, 
> although it may be negligible, by khugepaged.

A bit here and there is manageable.  There is, of course, some work to
be done there, but for now we're mainly concerned with a job that's
supposed to be confined to a cpuset spilling out and soaking up all the
memory on a machine.

Thanks for the help, David.  Much appreciated!

- Alex
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/