Re: [PATCH net] net: phy: Decrement phy_fixed_addr during unregister

On 06/24/2016 04:06 PM, Russell King - ARM Linux wrote:

On Fri, Jun 24, 2016 at 03:58:39PM -0700, Florian Fainelli wrote:

quoted

On 06/24/2016 03:55 PM, Russell King - ARM Linux wrote:

quoted

On Fri, Jun 24, 2016 at 03:44:11PM -0700, Florian Fainelli wrote:

quoted

If we have a system which uses fixed PHY devices and calls
fixed_phy_register() then fixed_phy_unregister() we can exhaust the
number of fixed PHYs available after a while, since we keep incrementing
the variable phy_fixed_addr, but we never decrement it.

This patch fixes that by decrementing phy_fixed_addr during
fixed_phy_del(), and in order to do that, we need to move the
phy_fixed_addr integer and its spinlock above that function.

Is this really a good idea?

In the sense that it is symetrical to the register code, probably.

quoted

What if we have two fixed phys register, and the first one is
unregistered and a new one subsequently registered?

First phy registered, gets address 0, phy_fixed_addr becomes 1.
Second phy registered, gets address 1, phy_fixed_addr becomes 2.
First phy is unregistered, phy_fixed_addr becomes 1.
Third phy registered, gets address 1, conflicts with the second phy.

Obviously not a good outcome.

What would you suggest we do instead? Would switching to IDA/IDR give us
better results for instance (I have not looked too closely yet)?

I would expect an IDA to be suitable, because the IDA would track which
indexes (==addresses) are currently in-use.

OK, thanks!

If you want to go further, using an IDR would allow fixed_mdio_read() to
find the right fixed_phy struct without needing to loop over fmb->phys.

Since I am targetting this as a bugfix, the switch to IDA seems more
appropriate to be backported, but yes, that's a good idea though.

Whether that's worth it or not depends if you have a large number of
fixed phys.  I suspect we're talking about small quantities here though.

Yes, at the moment we are limited to 32 PHYs maximum, just like a real
MDIO bus, which in some systems could actually be not enough, but then
you run into other problems, like the need to register more than a
single fixed MDIO bus driver to get a larger address space...
-- 
Florian