Re: [PATCH net-next v6 3/3] net: ethernet: ti: am65-cpsw: Add support for SERDES configuration

[Siddharth Vadapalli]

Hello Geert,

On 18/01/23 15:57, Geert Uytterhoeven wrote:
> Hi Siddarth,
> 
> On Wed, Jan 18, 2023 at 6:48 AM Siddharth Vadapalli <s-vadapalli@xxxxxx> wrote:
>> On 17/01/23 19:25, Geert Uytterhoeven wrote:
>>> On Wed, Jan 4, 2023 at 11:37 AM Siddharth Vadapalli <s-vadapalli@xxxxxx> wrote:
>>>> Use PHY framework APIs to initialize the SERDES PHY connected to CPSW MAC.
>>>>
>>>> Define the functions am65_cpsw_disable_phy(), am65_cpsw_enable_phy(),
>>>> am65_cpsw_disable_serdes_phy() and am65_cpsw_enable_serdes_phy().
>>>>
>>>> Add new member "serdes_phy" to struct "am65_cpsw_slave_data" to store the
>>>> SERDES PHY for each port, if it exists. Use it later while disabling the
>>>> SERDES PHY for each port.
>>>>
>>>> Power on and initialize the SerDes PHY in am65_cpsw_nuss_init_slave_ports()
>>>> by invoking am65_cpsw_enable_serdes_phy().
>>>>
>>>> Power off the SerDes PHY in am65_cpsw_nuss_remove() by invoking
>>>> am65_cpsw_disable_serdes_phy().
>>>>
>>>> Signed-off-by: Siddharth Vadapalli <s-vadapalli@xxxxxx>
>>>
>>> Thanks for your patch, which is now commit dab2b265dd23ef8f ("net:
>>> ethernet: ti: am65-cpsw: Add support for SERDES configuration")
>>> in net-next.
>>>
>>>> --- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c
>>>> +++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
> 
>>>> +static int am65_cpsw_init_serdes_phy(struct device *dev, struct device_node *port_np,
>>>> +                                    struct am65_cpsw_port *port)
>>>> +{
>>>> +       const char *name = "serdes-phy";
>>>> +       struct phy *phy;
>>>> +       int ret;
>>>> +
>>>> +       phy = devm_of_phy_get(dev, port_np, name);
>>>> +       if (PTR_ERR(phy) == -ENODEV)
>>>> +               return 0;
>>>> +
>>>> +       /* Serdes PHY exists. Store it. */
>>>
>>> "phy" may be a different error here (e.g. -EPROBE_DEFER)...
>>
>> The Serdes is automatically configured for multi-link protocol (Example: PCIe +
>> QSGMII) by the Serdes driver, due to which it is not necessary to invoke the
>> Serdes configuration via phy_init(). However, for single-link protocol (Example:
>> Serdes has to be configured only for SGMII), the Serdes driver doesn't configure
>> the Serdes unless requested. For this case, the am65-cpsw driver explicitly
>> invokes phy_init() for the Serdes to be configured, by looking up the optional
>> device-tree phy named "serdes-phy". For this reason, the above section of code
>> is actually emulating a non-existent "devm_of_phy_optional_get()". The
>> "devm_of_phy_optional_get()" function is similar to the
>> "devm_phy_optional_get()" function in the sense that the "serdes-phy" phy in the
>> device-tree is optional and it is not truly an error if the property isn't present.
> 
> Yeah, I noticed while adding devm_phy_optional_get(), and looking for
> possible users.
> See "[PATCH treewide 0/7] phy: Add devm_of_phy_optional_get() helper"
> https://lore.kernel.org/all/cover.1674036164.git.geert+renesas@xxxxxxxxx

Thank you for working on this.

> 
>> Thank you for pointing out that if the Serdes driver is built as a module and
>> the am65-cpsw driver runs first, then the "phy" returned for "serdes-phy" will
>> be "-EPROBE_DEFER".
>>
>>>
>>>> +       port->slave.serdes_phy = phy;
>>>> +
>>>> +       ret =  am65_cpsw_enable_phy(phy);
>>>
>>> ... so it will crash when dereferencing phy in phy_init().
>>>
>>> I think you want to add an extra check above:
>>>
>>>     if (IS_ERR(phy))
>>>             return PTR_ERR(phy);
>>
>> Please let me know if posting a "Fixes" patch for fixing this net-next commit is
>> the right process to address this.
> 
> I think it is, as devm_of_phy_optional_get() might not make it in time.

I posted the patch at:
https://lore.kernel.org/r/20230118112136.213061-1-s-vadapalli@xxxxxx

> 
>>>> @@ -1959,6 +2021,11 @@ static int am65_cpsw_nuss_init_slave_ports(struct am65_cpsw_common *common)
>>>
>>> Right out of context we have:
>>>
>>>                 port->slave.ifphy = devm_of_phy_get(dev, port_np, NULL);
>>>                 if (IS_ERR(port->slave.ifphy)) {
>>>                         ret = PTR_ERR(port->slave.ifphy);
>>>                         dev_err(dev, "%pOF error retrieving port phy: %d\n",
>>>                                 port_np, ret);
>>>
>>> So if there is only one PHY (named "serdes-phy") in DT, it will be
>>> used for both ifphy and serdes_phy. Is that intentional?
>>
>> The PHY corresponding to "ifphy" is meant to be the CPSW MAC's PHY and not the
>> Serdes PHY. The CPSW MAC's PHY is configured by the
>> drivers/phy/ti/phy-gmii-sel.c driver and this is NOT an optional PHY, unlike the
>> Serdes PHY. Therefore, it is assumed that the CPSW MAC's PHY is always provided
>> in the device-tree, while the Serdes PHY is optional, depending on whether the
>> Serdes is being configured for single-link protocol or multi-link protocol.
>> Please let me know if this appears to be an issue and I will fix it based on
>> your suggestion.
> 
> Hence this should be documented in the DT bindings. Please document
> there can be 1 or 2 phys, with an optional "phys-names" property,
> listing "ifphy" and "serdes-phy" (the DT people might request a rename).

I will work on this.

Regards,
Siddharth.