Re: [RFC PATCH] fpga: remove module reference counting from core components

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

On Thu, Nov 09, 2023 at 03:16:26PM +0800, Xu Yilun wrote:
> On Thu, Nov 09, 2023 at 06:27:24AM +0100, Greg Kroah-Hartman wrote:
> > On Thu, Nov 09, 2023 at 01:07:42PM +0800, Xu Yilun wrote:
> > > On Wed, Nov 08, 2023 at 05:20:53PM +0100, Greg Kroah-Hartman wrote:
> > > > On Wed, Nov 08, 2023 at 11:52:52PM +0800, Xu Yilun wrote:
> > > > > > >>
> > > > > > >> In fpga_region_get() / fpga_region_put(): call get_device() before
> > > > > > >> acquiring the mutex and put_device() after having released the mutex
> > > > > > >> to avoid races.
> > > > 
> > > > Why do you need another reference count with a lock?  You already have
> > > > that with the calls to get/put_device().
> > > 
> > > The low-level driver module could still be possibly unloaded at the same
> > > time, if so, when FPGA core run some callbacks provided by low-level driver
> > > module, its referenced page of code is unmapped...
> > 
> > Then something is designed wrong here, the unloading of the low-level
> > driver should remove the access to the device itself.  Perhaps fix that?
> 
> Actually the low-level driver module on its own has no way to garantee its
> own code page of callbacks not accessed. It *is* accessing its code page
> when it tries (to release) any protection.

It is not up to the low-level driver to do this, it's up to the code
that calls into it (i.e. the fpga core code) to handle the proper
reference counting.

> Core code must help, and something like file_operations.owner is an
> effective way.

Yes, that should be all that you need.

thanks,

greg k-h