Re: dentry cache leak? Re: rsync out of memory 2.6.8-rc2

[kladit@t-online.de (Klaus Dittrich)]

Nick Piggin wrote:

> Klaus Dittrich wrote:
>
>> Klaus Dittrich wrote:
>
>
>>> I did a test with a value of 500. echo 500 > 
>>> /proc/sys/vm/vfs_cache_pressure.
>>>
>>> The highest numbers a cat /proc/sys/fs/dentry-state then showed during
>>> a du -s were
>>> 780721  750505  45      0       0       0
>>>
>>> The system survied. No processes were killed.
>>>
>>> With vfs_cache_pressure=100 a cat /proc/sys/fs/dentry-state showed
>>> numbers of about 1090000 before processes got killed.
>>>
>>> Hope that helps to narrow the region to look for what has changed.
>>>
>> PS. Two concurrent du -s however "kernel: Out of Memory: Killed 
>> process .." *
>> *
>
>
> Your vfs_cache_pressure probably wants to be higher than 500. Make it 
> 10000.

*No problems when using 10000.
But I think of this as workaround only.

I have read the "Scaling dcache with RCU" article from linuxjournal.com
to get some insight how things should work. Pretty complicated.

I added some printk statements in dcache.c to see what functions
actually get called to shink a dcache during a du -s.
The only one I found is prune_dcache().
Which events should trigger the use of d_invalidate() ?


**
*