On Wed, 2023-10-11 at 08:43 +0200, Marc Kleine-Budde wrote:
> On 11.10.2023 00:21:31, Francois Romieu wrote:
> > Marc Kleine-Budde <mkl@xxxxxxxxxxxxxx> :
> > > The dm9000 takes the db->lock spin lock in dm9000_timeout() and calls
> > > into dm9000_init_dm9000(). For the DM9000B the PHY is reset with
> > > dm9000_phy_write(). That function again takes the db->lock spin lock,
> > > which results in a deadlock. For reference the backtrace:
> > [...]
> > > To workaround similar problem (take mutex inside spin lock ) , a
> > > "in_timeout" variable was added in 582379839bbd ("dm9000: avoid
> > > sleeping in dm9000_timeout callback"). Use this variable and not take
> > > the spin lock inside dm9000_phy_write() if in_timeout is true.
> > >
> > > Cc: stable@xxxxxxxxxxxxxxx
> > > Signed-off-by: Marc Kleine-Budde <mkl@xxxxxxxxxxxxxx>
> > > ---
> > > During the netdev watchdog handling the dm9000 driver takes the same
> > > spin lock twice. Avoid this by extending an existing workaround.
> > > ---
> >
> > I can review it but I can't really endorse it. :o)
> >
> > Extending ugly workaround in pre-2000 style device drivers...
> > I'd rather see the thing fixed if there is some real use for it.
>
> There definitely are still users of this drivers on modern kernels out
> there.
>
> I too don't like the feeling of wrapping more and more duct tape
> around existing drivers. How about moving the functionality to
> dm9000_phy_write_locked() and leave the locking in dm9000_phy_write().
> I will prepare a patch.
If you have the H/W handy to try some more invasive change, I'm
wondering if you could schedule a work from dm9000_timeout() and there
acquire all the needed locks.
Cheers,
Paolo
