Re: Problems with get_driver() and driver_attach() (and new_id too)

[Dmitry Torokhov <dmitry.torokhov@gmail.com>]

On Mon, Jan 09, 2012 at 11:37:57AM -0500, Alan Stern wrote:
> On Mon, 9 Jan 2012, Dmitry Torokhov wrote:
> 
> > > I don't think any of those calls actually accomplish anything, but it's
> > > hard to be certain.  Some of them appear to be futile attempts to
> > > prevent the driver from being unregistered or unloaded, others are
> > > there simply to drop the reference taken by driver_find().
> > > 
> > > In a few of them it's obvious that the driver can't be unregistered 
> > > while the critical section runs, but in the others I can't tell.  On 
> > > the other hand, if a critical section can race with unregistration 
> > > then the code is buggy now.
> > > 
> > > What do you think?
> > 
> > I think we need to audit them and decide on case-by-case basis. For
> > example drivers/s390/cio/device.c is completely nonsensical: it takes a
> > reference on a driver that is passed as argument before calling
> > driver_find_device(). But if passed driver was valid before we called
> > get_driver it won't become any more valid afterwards and it should not
> > disappear either.
> > 
> > drivers/s390/cio/ccwgroup.c - calls are useless;
> > 
> > Authors of drivers/net/phy/phy_device.c had their reservations:
> > 
> >         /* Make sure the driver is held.
> >          * XXX -- Is this correct? */
> >         drv = get_driver(phydev->dev.driver);
> > 
> > However it is in phydev_probe() and I hope our device core takes care of
> > not destroying drivers in the middle of binding to a device.
> 
> Yes, it does.  That one looks like a misunderstanding.  It calls
> get_driver during phy_probe and put_driver during phy_remove, which
> accomplishes nothing.
> 
> > drivers/ssb/main.c seems like needs some protection but does it
> > incorrectly as we do not wait for drivers to drop all references before
> > unloading modules.
> 
> Possibly it needs to be replaced with try_module_get.  I'll send out an 
> email to the maintainers of these drivers to see what they think.

No, I am not that try_module_get() [alone] is quite what is needed, as
strictly speaking driver lifetime does not need to be the same as module
lifetime. But maybe I am just splitting hair as all drivers are
statically initialized and are tied to their modules...

-- 
Dmitry