Re: [Gumstix-users] [PATCH]: uClibc dlclose() bug: solved. (well, not that much)

[Alexandre P. N. <al...@om...>]

ken staton escreveu:

>I tried this out on the python module import segfault  issue.
>Still there.
>  
>

Yes, it only solves one minor problem, I was fooled into thinking it 
would take care of it all because my binaries were already cached. In 
fact the problem this patch solves is hardly noticeable, most programs 
don't dlclose() anything except on shutdown.

Since it's a problem after all and the patch is included on uclibc's svn 
version, I recommend keeping it for the time being while we dig into the 
other problem.

I'm facing a hard time chasing the bug by lack of information; It seems 
that as this problem manifests rarely, people on the arm-linux mailing 
list are a bit skeptical about it's very existance. There's some 
discussion going on the main linux kernel list, but I couldn't get up 
with it yet.

It's probably a cache handling issue, because it's very unlikely that 
the page fault handling would be that buggy. Other (less likely) 
possibilities I was considering was some jffs2 fs page load atomicity 
issue or even another uclibc's loader bug. The latter was supported 
because while tracing it in userspace with gdb, I could verify that the 
address reported by the kernel as the faulting page (when booting with 
parameter user_debug=0x1f) happened on an definitively unmapped area 
(also verifiable on /proc/<pid>/maps).

I don't even feel I'm going on the right direction right now. :-)

- Alexandre