Re: Memory tables corruption - kernel v2.4.33-rc1

[Willy Tarreau <w@1wt.eu>]

Hello,

On Mon, Jul 10, 2006 at 08:39:38PM -0800, Tony Borras wrote:
> Please fwd to 2.4.33 Team.
> 
> One of my customers reported in tests that, after some runtime,
> his 59MB ram drive suddenly, randomly shrinks.
> 
> Same customer s/w has been running fine with 2.4.31 kernel.
> 
> 
> Report follows:
> 
> I have two Advantech (i486 Geode's w/64MB ram, ~49MB usable, as
> configured in the kernel make) test units where Dosemu crashed
> this weekend. For some reason the Via unit was fine.
> 
> Here is what happened:
> 
> 1. Linux was O.K., dosemu 1.3.2 had shut down.
> It appears both units were out of RAM.
> 
> 2. The first unit showed the RAM drive size to be
> 14661 1k blocks using the df command. The 1K blocks
> should be 47595, as they were after boot. The other unit
> showed the correct value of 1K blocks=47595. Even after
> restarting the Advantech, the 1K blocks would not go back to
> normal, so my software cannot even load.
> 
> 3. The second Advantech unit showed correct values with the df
> command, but when I used the du -cbs command, I found the /tmp
> directory had only 110796 bytes, instead of the normal ~15 MB.
> When I restarted Linux, all the RAM came back, and the
> unit started working again.
> 
>                               ---------
> 
> Thanks, I will check further with this customer.

2.4.33-rc1 has a bug in vfs_unlink(). Probably you've been hitting
it when removing lots of files from /tmp which might not have been
really freed. -rc2 has this fixed. You should try it instead. Other
people reported kernel panics on -rc1 with NFS.

> Tony Borras

Regards,
Willy