Re: [PATCH net-next] net: phy: mdio-boardinfo: Allow recursive mdiobus_register()

[Andrew Lunn <andrew@lunn.ch>]

On Wed, Apr 18, 2018 at 05:14:36PM -0700, Florian Fainelli wrote:
> On 04/18/2018 05:00 PM, Andrew Lunn wrote:
> > mdiobus_register will search for any mdiobus board info registered for
> > the bus being registered. If found, it will probe devices on the bus.
> > That device, if for example it is an ethernet switch, may then try to
> > register an mdio bus. Thus we need to allow recursive calls to
> > mdiobus_register.
> > 
> > Holding the mdio_board_lock will cause a deadlock during this
> > recursion. Release the lock and use list_for_each_entry_safe.
> > 
> > Signed-off-by: Andrew Lunn <andrew@lunn.ch>
> > ---
> >  drivers/net/phy/mdio-boardinfo.c | 5 ++++-
> >  1 file changed, 4 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/net/phy/mdio-boardinfo.c b/drivers/net/phy/mdio-boardinfo.c
> > index 1861f387820d..863496fa5d13 100644
> > --- a/drivers/net/phy/mdio-boardinfo.c
> > +++ b/drivers/net/phy/mdio-boardinfo.c
> > @@ -30,17 +30,20 @@ void mdiobus_setup_mdiodev_from_board_info(struct mii_bus *bus,
> >  					    struct mdio_board_info *bi))
> >  {
> >  	struct mdio_board_entry *be;
> > +	struct mdio_board_entry *tmp;
> >  	struct mdio_board_info *bi;
> >  	int ret;
> >  
> >  	mutex_lock(&mdio_board_lock);
> 
> Don't you need to drop that lock here?
> 
> > -	list_for_each_entry(be, &mdio_board_list, list) {
> > +	list_for_each_entry_safe(be, tmp, &mdio_board_list, list) {
> >  		bi = &be->board_info;
> >  
> >  		if (strcmp(bus->id, bi->bus_id))
> >  			continue;
> >  
> > +		mutex_unlock(&mdio_board_lock);
> >  		ret = cb(bus, bi);
> > +		mutex_lock(&mdio_board_lock);
> >  		if (ret)
> >  			continue;
> >  
> 
> And conversely drop the unlock from the end of this function?

No. The recursion happens inside the ret = cb(bus, bi). We need the
lock to be available during that. The lock itself is protecting the
list, so we need to hold the lock while using the list.

> Also, would you rather target "net" for this change since this appears
> to be a bug fix?

As far as i know, there is no in tree driver which can trigger
this. It requires the use of a switch instantiated using mdio_board
info, and the switch then needs to add another mdio bus in its probe
function. The only in tree code doing anything like this is dsa_loop,
and it does not register an mdio bus.

However, later this cycle i plan to add support for Zodiac's SCU3, and
it does trigger this deadlock. So net-next is sufficient for my.

   Andrew