Re: [PATCH] Fix 2.5.3pre reiserfs BUG() at boot time

[Andreas Dilger <adilger@turbolabs.com>]

On Jan 25, 2002  10:08 -0800, Linus Torvalds wrote:
> I would prefer instead just avoiding the copy altogether, and just save
> the name pointer - with no length restrictions.
> 
> Right now the code has the comment
> 
>    /* Copy name over so we don't have problems with unloaded modules */

Yes, I put that in.

> but that was written before "kmem_cache_destroy()" existed, and we should
> long ago have fixed any modules that don't properly destroy their caches
> when they exit (and yes, I know the difference between "should" and "did",
> but that's not an excuse for a bad interface).

The problem is that if, for some reason, the cache is NOT empty when you
call kmem_cache_destroy(), it will not be freed, but the module exits
anyways.  Then, any access to /proc/slabinfo will OOPS.

Yes, code should be written correctly so that its slab is empty when it
exits, but I'd rather have a _bit_ of safety here so that you can at
least check slabinfo when you get a kernel message "slab is not empty"
(or whatever it is) so you can at least try and investigate the problem.

The other alternative is to BUG with enough information to figure out
the status of this cache if you try to free a non-empty cache.  At
least then you would get some data at the time the real problem happens
as opposed to killing some random process later that tries to read
slabinfo.

Cheers, Andreas
--
Andreas Dilger
http://sourceforge.net/projects/ext2resize/
http://www-mddsp.enel.ucalgary.ca/People/adilger/