Re: 2.6.8.1 mempool subsystem sickness

[Nick Piggin <nickpiggin@yahoo.com.au>]

Jeff V. Merkey wrote:
> Nick Piggin wrote:

>> OK, this is against 2.6.9-rc2. Let me know how you go. Thanks
>>
>>  
>>
> 
> Nick,
> 
> The problem is corrected with this patch.  I am running with 3GB of 
> kernel memory
> and 1GB user space with the userspace splitting patch with very heavy 
> swapping
> and user space app activity and no failed allocations.  This patch 
> should be rolled
> into 2.6.9-rc2 since it fixes the problem.  With standard 3GB User/1GB 
> kernel
> address space, it also fixes the problems with X server running out of 
> memory
> and the apps crashing.
> 

Hi Jeff,
Thanks, that is very cool. The memory problems you're seeing aren't
actually a regression (it's always been like that), and I still haven't
got hold of some gigabit networking hardware to test it thoroughly, so
as such so it may be difficult to get this into 2.6.9. Hopefully soon
though.

I can provide you (or anyone) with up to date patches on request though,
so just let me know.

> Jeff
> 
> Here's the stats from the test of the patch against 2.6.8-rc2 with the 
> patch applied
> 
> 

Scanning stats look good at a quick glance. kswapd doesn't seem to be
going crazy.

However,
size-65536         32834  32834  65536    1   16

This slab entry is taking up about 2GB of unreclaimable memory (order-4,
no less). This must be a leak... does the number continue to rise as
your system runs?