Re: slab: setup allocators earlier in the boot sequence

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Fri, 2009-06-12 at 16:54 +0300, Pekka Enberg wrote:
> Hi Christoph,
> 
> On Fri, 2009-06-12 at 09:49 -0400, Christoph Lameter wrote:
> > Best thing to do is to recognize the fact that we are still in early boot
> > in the allocators. Derived allocators (such as slab and vmalloc) mask bits
> > using GFP_RECLAIM_MASK and when doing allocations through the page
> > allocator. You could make GFP_RECLAIM_MASK a variable. During boot
> > __GFP_WAIT would not be set in GFP_RECLAIM_MASK.
> 
> Ben's patch does something like that and I have patches that do that
> floating around too.
> 
> The problem here is that it's not enough that we make GFP_RECLAIM_MASK a
> variable. There are various _debugging checks_ that happen much earlier
> than that. We need to mask out those too which adds overhead to
> kmalloc() fastpath, for example.

Hrm... I though I stuck my masking before the lockdep tests but maybe I
missed some...

Again, I'm not saying my patch is the best way to solve it. My point is
more that the callers shouldn't have to bother. (And thus the WARN_ON
isn't right :-)

Cheers,
Ben.