SKAS patch (was: Re: [uml-devel] uml memory leak)

[BlaisorBlade <bla...@ya...>]

Alle 17:08, sabato 22 maggio 2004, Nicolas ha scritto:
> Hello,
>
> Here is my hostos config :
>
> fresh 2.6.6 kernel with :
>
> uml-patch-2.6.6-02
> patch-2.6.6-v2
> host+Uml-skas3-2.6.6-v1.patch
>
> I noticed a memory leak each time starting/stopping
> an uml os. (approx 7mo).

I resume his datas:
Before:
size-4096          17898  17898   4096    1    1 : tunables   24   12    0 : 
slabdata  17898  17898      0
size-32            20157  20679     32  113    1 : tunables  120   60    0 : 
slabdata    183    183      0

After starting and stopping 10 UMLs:
size-4096          37539  37539   4096    1    1 : tunables   24   12    0 : 
slabdata  37539  37539      0
size-32            39791  40228     32  113    1 : tunables  120   60    0 : 
slabdata    356    356      0

> I don't know what it means, but size-32 and size-4096 are increasing ...

I can confirm it on 2.4.26 with the SKAS patch I prepared (which is VERY 
similar to the 2.6 one), and preemption + O(1) scheduler (actually, the whole 
-lck tree).

Sadly I did not save my /proc/meminfo before rebooting, but I just found about 
256M leaked by the host, after running multiple UML in skas mode. I expect 
this not to happen in TT mode (this is probably a bug in the SKAS patch).

There were about 64*1000 objects inside size-4096 (i.e. each one wasting 4k), 
and notably about the same number of objects inside size-32 (i.e. we leak 
this things together). Btw, I went down to runlevel 1 and those objects were 
still there. So it's a kernel leak. It is not a mm_struct leak, so I don't 
understand where it comes from.
-- 
Paolo Giarrusso, aka Blaisorblade
Linux registered user n. 292729