Re: [PATCH v1 net-next] net: phy: mdio_bus: make mdiobus_scan also cover PHY that only talks C45

From: Florian Fainelli
Date: Wed Aug 28 2019 - 13:00:09 EST


On 8/28/19 8:41 AM, Ong, Boon Leong wrote:
>> On Tue, Aug 27, 2019 at 03:23:34PM +0000, Voon, Weifeng wrote:
>>>>>> Make mdiobus_scan() to try harder to look for any PHY that only
>>>> talks C45.
>>>>> If you are not using Device Tree or ACPI, and you are letting the MDIO
>>>>> bus be scanned, it sounds like there should be a way for you to
>>>>> provide a hint as to which addresses should be scanned (that's
>>>>> mii_bus::phy_mask) and possibly enhance that with a mask of possible
>>>>> C45 devices?
>>>>
>>>> Yes, i don't like this unconditional c45 scanning. A lot of MDIO bus
>>>> drivers don't look for the MII_ADDR_C45. They are going to do a C22
>>>> transfer, and maybe not mask out the MII_ADDR_C45 from reg, causing an
>>>> invalid register write. Bad things can then happen.
>>>>
>>>> With DT and ACPI, we have an explicit indication that C45 should be used,
>>>> so we know on this platform C45 is safe to use. We need something
>>>> similar when not using DT or ACPI.
>>>>
>>>> Andrew
>>>
>>> Florian and Andrew,
>>> The mdio c22 is using the start-of-frame ST=01 while mdio c45 is using ST=00
>>> as identifier. So mdio c22 device will not response to mdio c45 protocol.
>>> As in IEEE 802.1ae-2002 Annex 45A.3 mention that:
>>> " Even though the Clause 45 MDIO frames using the ST=00 frame code
>>> will also be driven on to the Clause 22 MII Management interface,
>>> the Clause 22 PHYs will ignore the frames. "
>>>
>>> Hence, I am not seeing any concern that the c45 scanning will mess up with
>>> c22 devices.
>>
>> Hi Voon
>>
>> Take for example mdio-hisi-femac.c
>>
>> static int hisi_femac_mdio_read(struct mii_bus *bus, int mii_id, int regnum)
>> {
>> struct hisi_femac_mdio_data *data = bus->priv;
>> int ret;
>>
>> ret = hisi_femac_mdio_wait_ready(data);
>> if (ret)
>> return ret;
>>
>> writel((mii_id << BIT_PHY_ADDR_OFFSET) | regnum,
>> data->membase + MDIO_RWCTRL);
>>
>>
>> There is no check here for MII_ADDR_C45. So it will perform a C22
>> transfer. And regnum will still have MII_ADDR_C45 in it, so the
>> writel() is going to set bit 30, since #define MII_ADDR_C45
>> (1<<30). What happens on this hardware under these conditions?
>>
>> You cannot unconditionally ask an MDIO driver to do a C45
>> transfer. Some drivers are going to do bad things.
>
> Andrew & Florian, thanks for your review on this patch and insights on it.
> We will look into the implementation as suggested as follow.
>
> - for each bit clear in mii_bus::phy_mask, scan it as C22
> - for each bit clear in mii_bus::phy_c45_mask, scan it as C45
>
> We will work on this and resubmit soonest.

Sounds good. If you do not need to scan the MDIO bus, another approach
is to call get_phy_device() by passing the is_c45 boolean to true in
order to connect directly to a C45 device for which you already know the
address.

Assuming this is done for the stmmac PCI changes that you have
submitted, and that those cards have a fixed set of addresses for their
PHYs, maybe scanning the bus is overkill?
--
Florian