Re: [patch 0/6] mm: alloc_percpu and bigrefs

[Dipankar Sarma <dipankar@in.ibm.com>]

On Fri, Sep 23, 2005 at 06:11:30PM +1000, Rusty Russell wrote:
> On Fri, 2005-09-23 at 00:17 -0700, David S. Miller wrote:
> > I'm still against expanding these networking datastructures with
> > bigrefs just for this stuff.  Some people have per-cpu and per-node on
> > the brain, and it's starting to bloat things up a little bit too much.
> 
> I think for net devices it actually makes sense; most of the time we are
> not trying to remove them, so the refcounting is simply overhead.  We
> also don't alloc and free them very often.  The size issue is not really
> an issue since we only map for each CPU, and even better: if a bigref
> allocation can't get per-cpu data it just degrades beautifully into a
> test and an atomic.

I agree, given that it is for a much smaller number of refcounted
objects.

> 
> Now, that said, I wanted (and wrote, way back when) a far simpler
> allocator which only worked for GFP_KERNEL and used the same
> __per_cpu_offset[] to fixup dynamic per-cpu ptrs as static ones.  Maybe
> not as "complete" as this one, but maybe less offensive.

The GFP_ATOMIC support stuff is needed only for dst entries. However
using per-cpu refcounters in such objects like dentries and dst entries
are problematic and that is why I hadn't tried changing those.
Some of the earlier versions of the allocator were simpler and I think
we need to roll this back and do some more analysis. No GFP_ATOMIC
support, no early use. I haven't got around to look at this 
for a while, but I will try.

Thanks
Dipankar