Re: PCI driver module unload race?

[Oliver Neukum <oliver@neukum.name>]

Am Dienstag, 11. März 2003 02:15 schrieb Greg KH:
> On Tue, Mar 11, 2003 at 02:04:20AM +0100, Roman Zippel wrote:
> > On Mon, 10 Mar 2003, Greg KH wrote:
> > > > It seems that the semaphore in bus_add_device() makes this
> > > > unnecessary.
> > >
> > > Hm, yes.  I think you are correct.
> > >
> > > So this patch is not needed, and the struct module * can be ripped out
> > > of struct usb_driver too :)
> >
> > I think it's not easy. I haven't studied the code completely yet, but
> > e.g. when you attach a device to a driver you also have to get a
> > reference to the driver.
>
> You get a link to the driver, but you can't increment the module count
> of the driver at that time, as we have to be able to remove a module
> somehow :)

That is simple. Export a generic way to disconnect a driver from a device.

> > I think there are more interesting races, e.g. when you create a sysfs
> > symlink, that symlink might also have references to a module.
>
> Yeah, I still think there are some nasty issues with regards to being in
> a sysfs directory, with a open file handle, and the module is removed.
> But I haven't checked stuff like that in a while.
>
> CONFIG_MODULE_UNLOAD, just say no.

That is taking the easy way out.

	Regards
		Oliver