Re: 2.6.14, udev: unknown symbols for ehci_hcd

[Greg KH <greg@kroah.com>]

On Mon, Nov 07, 2005 at 08:20:15PM +0100, Pozsar Balazs wrote:
> On Mon, Nov 07, 2005 at 11:12:14AM -0800, Greg KH wrote:
> > On Mon, Nov 07, 2005 at 08:07:38PM +0100, Pozsar Balazs wrote:
> > > On Mon, Nov 07, 2005 at 09:31:57AM -0800, Greg KH wrote:
> > > > On Mon, Nov 07, 2005 at 12:33:29PM +0100, Marco d'Itri wrote:
> > > > > On Nov 06, Greg KH <greg@kroah.com> wrote:
> > > > > 
> > > > > > This seems to be a Debian issue for some odd reason, I suggest filing a
> > > > > > bug against the udev package (or just tagging onto the existing bug for
> > > > > > this problem, I've seen it in there already...)
> > > > > The reason this is usually seen only on Debian systems is that I am the
> > > > > first one who shipped an udev package which runs many parallel modprobe
> > > > > commands, but this is a genuine kernel/modprobe bug.
> > > > 
> > > > I'm pretty sure OpenSuSE 10.0 does the same thing, and I don't think
> > > > anyone has reported the same kind of bugs there.  Makes me wonder what
> > > > is really happening here...
> > > 
> > > If module A depends on module B, and "modprobe A" and "modprobe B" are 
> > > run parallel
> > 
> > Why would they be run in parallel?  modprobe doesn't do this, why would
> > you?
> 
> On my machine it happened while loading pcmcia modules. It was two 
> modprobes running in parallel (invoked by udev), not 1 modprobe loading 
> modules in parallel.

Ah, ok, that explains how it is happening, thanks.

> > > , there is time window when module B is already listed in 
> > > /proc/modules, but not completely loaded/initialized, it is in the state 
> > > "Loading". At this point "modprobe A" checks /proc/modules if module B 
> > > is already loaded, but it does not take into account that it is in the 
> > > state "Loading" and not yet "Live". So it tries to load module A, but it 
> > > fails, because there are missing symbols because module A did not 
> > > register them yet.
> > 
> > Sounds like a locking issue within the module core.  I thought we could
> > only load one at a time, otherwise we have other races within the
> > kernel.
> 
> That's what I thought too, but this is not the case it seems.

Um, why not?  Rusty?

thanks,

greg k-h