Re: Module init call vs symbols exporting race?

[Jan Glauber <jang@linux.vnet.ibm.com>]

On Thu, 2007-11-08 at 13:10 +1100, Rusty Russell wrote:
> On Wednesday 07 November 2007 21:01:30 Jan Glauber wrote:
> > Hi Rusty,
> >
> > I've seen a symbol-resolving race on s390. The qeth module uses symbols
> > from qdio and although the loading order seems correct and the qdio
> > symbols should be available the following error appears:
> >
> > qdio: loading QDIO base support version 2
> > qeth: Unknown symbol qdio_synchronize
> 
> Looks like qdio does something which triggers qeth to load, but of course qdio 
> isn't finished initializing yet so its symbols aren't available.
> 
> It's not obvious what's triggering the load, but you could probably find it by 
> using printk's through qdio.c's init_QDIO().

Digging through the module loader I found what triggers the error...

CPU0 (sys_init_module for qdio)			CPU1 (sys_init_module for qeth)

mutex_lock()
-> load_module()
     mod->state = COMING
mutex_unlock()
						mutex_lock()
init().......takes some time			load_module()
						-> resolve_symbols()
						   -> use_module()
						      -> stong_try_module_get() bails out
							 because state == COMING
						-> simplify_symbols() complains with the warning

mutex_lock()
mod->state = LIVE
mutex_unlock()

So is it correct that sys_init_module() is called for qeth even if qdio is
not yet in MODULE_STATE_LIVE?

-jang