Re: -mm merge plans for 2.6.26 (memcgroup)

[Hugh Dickins <hugh@veritas.com>]

On Sun, 20 Apr 2008, Andrew Morton wrote:
> 
> These can be found at
> ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.25/2.6.25-mm1

A couple of comments on memcgroup patches in your
high-priority initial section (I've not studied further yet)
> Merge, and backport to 2.6.25.x

>  disable-the-memory-controller-by-default-v3.patch
>  disable-the-memory-controller-by-default-v3-fix.patch

If those are to go in, then the sooner the better, yes.

But though I argued for cgroup_disable=memory (or some such),
I think myself that taking it even further now (requiring an
additional cgroup_enable=memory at boottime to get the memcg
stuff you chose with CGROUP_MEM_RES_CTLR=y at build time) is
confusing overkill, just messing around.

Others think differently.  A compromise would be to improve the
helptext for CGROUP_MEM_RES_CTLR (some of it is presently nonsense,
isn't it?  Certainly there's a significant overhead, but it's the
32-bit struct page not the 64-bit which then suffers from crossing
cacheline boundaries).  Not much point in mentioning
cgroup_disable=memory if those patches go in, but needs to say
cgroup_enable=memory bootoption also needed.

>  memcgroup-check-and-initialize-page-cgroup-in-memmap_init_zone.patch

No, it was a good find from Shi, but you were right to think the patch
fishy, and Kame put in lots of work (thank you!) to identify the actual
culprit: he and Mel are discussing what the actual fix should be; and
we might want to choose a different fix for stable than for 2.6.26.

I think you should drop that memmap_init_zone patch: the cgroup
pointer is not the only field we assume is zeroed, both flags and
mapping can cause trouble if they were not originally zeroed.
Re-zero the whole struct page?  No, far better to fix the
root of the corruption, that Kame and Mel are working on.

Hugh