Re: page allocation failure & sk98lin

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

Serban Simu wrote:
> Thank you, Nick. Just wanted to mention that while I understand that we 
> recover from this allocation failure (and also I don't mind the stack 
> printouts), about 20% of my incoming network traffic (600-700 Mbps) 
> seems to be dropped in the process.

Yeah that is expected - so I guess it isn't exactly 'harmless' if
performance is critical.

> Does the memory manager have to 
> spend a considerable amount of time to recover?
> 
> I will have a look at the -mm fixes, thanks for the idea.
> 

The relevant patch is this one which is now merged into 2.6.

http://linux.bkbits.net:8080/linux-2.5/cset@412b8828ClkE2ZwNwGQ02aYoMwb7-A?nav=index.html|ChangeSet@-1d

It would be nice if you can test that. It will give GFP_ATOMIC allocators
a larger buffer between starting memory reclaim, and failing their allocations.

However if it is a production system and you can't test patches, Marcelo
pointed out that you should be able to work around the problem by increasing
/proc/sys/vm/min_free_kbytes. Then maybe you could try 2.6.9 with min_free_kbytes
back to its default setting :)