Re: ioctl32_unregister_conversion & modules

[Arnd Bergmann <arnd@arndb.de>]

Pavel Machek wrote:

> ...what is the problem?
> 
> It seems that function pointers into modules do not need any special
> treatmeant [I *know* there was talk about this on l-k; but I can't
> find anything in Documentation/]:
> 
>                 if (!capable(CAP_SYS_ADMIN))
>                         return -EACCES;
>                 if (disk->fops->ioctl) {
>                         ret = disk->fops->ioctl(inode, file, cmd, arg);
>                         if (ret != -EINVAL)
>                                 return ret;
>                 }

This is protected against unload by the reference counting done in 
open()/release(). ->ioctl() can be called only for open devices,
so you know the ioctl handler is not getting unloaded while it
is running.
 
> So... what's the problem with {un}register_ioctl32_conversion being
> called from module_init/module_exit? Drivers in the tree do it
> already...

The problem is that when the conversion handler is called, the reference
counting is only done for the module listed as ->owner in the
file operations. For example in the patch you submitted to add 
register_ioctl32_conversion() to drivers/serial/core.c I see nothing
stopping you from unloading core.ko while the handler is running
on a device owned by drivers/char/cyclades.c or any other serial driver.
It does not even have to be run on a serial driver, a user might try
to do ioctl(TIOCGSERIAL, ...) on a regular file...

        Arnd <><