Re: dentry_cache using up all my zone normal memory

["Christopher Friesen" <cfriesen@nortel.com>]

Marcelo wrote:

> How can this be reproduced? (point to a URL if you already explained
> that in detail).

I mentioned this earlier in the thread.  I'm running 2.6.14-rc4 on a 
pentium-M based atca blade with a bit over 3GB of memory.  When I run 
the "rename14" test from LTP with /tmp mounted on tmpfs, the system runs 
out of zone normal memory and the OOM killer starts killing processes.

If I have /tmp mounted nfs, the problem doesn't occur.  If I use the 
boot args to limit the memory to 896MB the problem doesn't occur.  If I 
run the testcase on a dual Xeon with multiple gigs of memory, the 
problem doesn't occur.

> Someone else on the thread said you had zillions of file descriptors
> open?

This does not appear to be the case.  The testcase has two threads.

thread 1 loops doing the following:
fd = creat("./rename14", 0666);
unlink("./rename14");
close(fd);

thread 2 loops doing:
rename("./rename14", "./rename14xyz");

> Need to figure out they can't be freed. The kernel is surely trying it
> (also a problem if it is not trying). How does the "slabs_scanned" field
> of /proc/vmstats looks like?

That's something I haven't checked...will have to get back to you.

> Bharata maintains a patch to record additional statistics, haven't 
> you tried it already?

Already tried.  You should be able to find results earlier in this thread.

Chris