module load/unload race protection?

module load/unload race protection?

[Mikael Pettersson]

Does the kernel's module loader (kernel/module.c, not kmod)
protect adequately against concurrent load/load or load/unload
requests? The question applies to both 2.2 and 2.4 kernels.

I'm trying to track down a problem where a user using a
RedHat 2.2.17-14 SMP kernel managed to trigger a situation where
a driver module had been unloaded while still being in use
(as in "the kernel has pointers into it", not USE_COUNT != 0).
I'm reviewing the driver's internal INC/DEC_USE_COUNT usage,
but so far I've not found any obvious errors.

/Mikael

Re: module load/unload race protection?

[Maneesh Soni]

On Sat, Apr 14, 2001 at 12:26:40PM +0200, Mikael Pettersson wrote:
> Does the kernel's module loader (kernel/module.c, not kmod)
> protect adequately against concurrent load/load or load/unload
> requests? The question applies to both 2.2 and 2.4 kernels.
> 
> I'm trying to track down a problem where a user using a
> RedHat 2.2.17-14 SMP kernel managed to trigger a situation where
> a driver module had been unloaded while still being in use
> (as in "the kernel has pointers into it", not USE_COUNT != 0).
> I'm reviewing the driver's internal INC/DEC_USE_COUNT usage,
> but so far I've not found any obvious errors.
> 
> /Mikael

There are race conditions while module unloading, present in both 2.2 and 2.4
kernels. Efforts are being made in handling these races by using
read-copy-update mechanism (http://lse.sourceforge.net/locking/rclock.html)
and by using synchronize_kernel() (from Rusty Russell, Keith Owens).
 
 
Regards,
Maneesh                                                                         

-- 
Maneesh Soni <smaneesh@in.ibm.com>
http://lse.sourceforge.net/locking/rclock.html
IBM Linux Technology Center,
IBM Software Lab, Bangalore, India