[RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Hugh Dickins]

On Mon, 29 Oct 2007, Balbir Singh wrote:
> On Mon, Oct 29, 2007 at 01:57:40AM +0530, Balbir Singh wrote:
> Hugh Dickins wrote:
> 
> [snip]
>  
> > Without your mem_cgroup mods in mm/swap_state.c, unuse_pte makes
> > the right assignments (I believe).  But I find that swapout (using
> > 600M in a 512M machine) from a 200M cgroup quickly OOMs, whereas
> > it behaves correctly with your mm/swap_state.c.
> > 
> 
> On my UML setup, I booted the UML instance with 512M of memory and
> used the swapout program that you shared. I tried two things
> 
> 
> 1. Ran swapout without any changes. The program ran well without
>    any OOM condition occuring, lot of reclaim occured.
> 2. Ran swapout with the changes to mm/swap_state.c removed (diff below)
>    and I still did not see any OOM. The reclaim count was much lesser
>    since swap cache did not get accounted back to the cgroup from
>    which pages were being evicted.
> 
> I am not sure why I don't see the OOM that you see, still trying. May be
> I missing something obvious at this late hour in the night :-)

I reconfirm that I do see those OOMs.  I'll have to try harder to
analyze how they come about: I sure don't expect you to debug a
problem you cannot reproduce.  But what happens if you try it
native rather than using UML?

Hugh