Re: [PATCH net-next 2/2] net: ethernet: ti: am65-cpsw: Enable USXGMII mode for J784S4 CPSW9G

[Siddharth Vadapalli]

On 03/04/23 15:27, Russell King (Oracle) wrote:
> On Mon, Apr 03, 2023 at 03:19:24PM +0530, Siddharth Vadapalli wrote:
>>
>>
>> On 03/04/23 14:29, Russell King (Oracle) wrote:
>>> On Mon, Apr 03, 2023 at 02:11:08PM +0530, Siddharth Vadapalli wrote:
>>>>
>>>>
>>>> On 03/04/23 14:02, Russell King (Oracle) wrote:
>>>>> On Mon, Apr 03, 2023 at 11:57:21AM +0530, Siddharth Vadapalli wrote:
>>>>>> Hello Russell,
>>>>>>
>>>>>> On 31/03/23 19:16, Siddharth Vadapalli wrote:
>>>>>>>
>>>>>>>
>>>>>>> On 31-03-2023 16:42, Russell King (Oracle) wrote:
>>>>>>>> On Fri, Mar 31, 2023 at 04:23:16PM +0530, Siddharth Vadapalli wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On 31/03/23 15:16, Russell King (Oracle) wrote:
>>>>>>>>>> On Fri, Mar 31, 2023 at 02:55:56PM +0530, Siddharth Vadapalli wrote:
>>>>>>>>>>> Russell,
>>>>>>>>>>>
>>>>>>>>>>> On 31/03/23 13:54, Russell King (Oracle) wrote:
>>>>>>>>>>>> On Fri, Mar 31, 2023 at 01:35:10PM +0530, Siddharth Vadapalli wrote:
>>>>>>>>>>>>> Hello Russell,
>>>>>>>>>>>>>
>>>>>>>>>>>>> Thank you for reviewing the patch.
>>>>>>>>>>>>>
>>>>>>>>>>>>> On 31/03/23 13:27, Russell King (Oracle) wrote:
>>>>>>>>>>>>>> On Fri, Mar 31, 2023 at 12:21:10PM +0530, Siddharth Vadapalli wrote:
>>>>>>>>>>>>>>> TI's J784S4 SoC supports USXGMII mode. Add USXGMII mode to the
>>>>>>>>>>>>>>> extra_modes member of the J784S4 SoC data. Additionally, configure the
>>>>>>>>>>>>>>> MAC Control register for supporting USXGMII mode. Also, for USXGMII
>>>>>>>>>>>>>>> mode, include MAC_5000FD in the "mac_capabilities" member of struct
>>>>>>>>>>>>>>> "phylink_config".
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> I don't think TI "get" phylink at all...
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
>>>>>>>>>>>>>>> index 4b4d06199b45..ab33e6fe5b1a 100644
>>>>>>>>>>>>>>> --- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c
>>>>>>>>>>>>>>> +++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
>>>>>>>>>>>>>>> @@ -1555,6 +1555,8 @@ static void am65_cpsw_nuss_mac_link_up(struct phylink_config *config, struct phy
>>>>>>>>>>>>>>>  		mac_control |= CPSW_SL_CTL_GIG;
>>>>>>>>>>>>>>>  	if (interface == PHY_INTERFACE_MODE_SGMII)
>>>>>>>>>>>>>>>  		mac_control |= CPSW_SL_CTL_EXT_EN;
>>>>>>>>>>>>>>> +	if (interface == PHY_INTERFACE_MODE_USXGMII)
>>>>>>>>>>>>>>> +		mac_control |= CPSW_SL_CTL_XGIG | CPSW_SL_CTL_XGMII_EN;
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> The configuration of the interface mode should *not* happen in
>>>>>>>>>>>>>> mac_link_up(), but should happen in e.g. mac_config().
>>>>>>>>>>>>>
>>>>>>>>>>>>> I will move all the interface mode associated configurations to mac_config() in
>>>>>>>>>>>>> the v2 series.
>>>>>>>>>>>>
>>>>>>>>>>>> Looking at the whole of mac_link_up(), could you please describe what
>>>>>>>>>>>> effect these bits are having:
>>>>>>>>>>>>
>>>>>>>>>>>> 	CPSW_SL_CTL_GIG
>>>>>>>>>>>> 	CPSW_SL_CTL_EXT_EN
>>>>>>>>>>>> 	CPSW_SL_CTL_IFCTL_A
>>>>>>>>>>>
>>>>>>>>>>> CPSW_SL_CTL_GIG corresponds to enabling Gigabit mode (full duplex only).
>>>>>>>>>>> CPSW_SL_CTL_EXT_EN when set enables in-band mode of operation and when cleared
>>>>>>>>>>> enables forced mode of operation.
>>>>>>>>>>> CPSW_SL_CTL_IFCTL_A is used to set the RMII link speed (0=10 mbps, 1=100 mbps).
>>>>>>>>>>
>>>>>>>>>> Okay, so I would do in mac_link_up():
>>>>>>>>>>
>>>>>>>>>> 	/* RMII needs to be manually configured for 10/100Mbps */
>>>>>>>>>> 	if (interface == PHY_INTERFACE_MODE_RMII && speed == SPEED_100)
>>>>>>>>>> 		mac_control |= CPSW_SL_CTL_IFCTL_A;
>>>>>>>>>>
>>>>>>>>>> 	if (speed == SPEED_1000)
>>>>>>>>>> 		mac_control |= CPSW_SL_CTL_GIG;
>>>>>>>>>> 	if (duplex)
>>>>>>>>>> 		mac_control |= CPSW_SL_CTL_FULLDUPLEX;
>>>>>>>>>>
>>>>>>>>>> I would also make mac_link_up() do a read-modify-write operation to
>>>>>>>>>> only affect the bits that it is changing.
>>>>>>>>>
>>>>>>>>> This is the current implementation except for the SGMII mode associated
>>>>>>>>> operation that I had recently added. I will fix that. Also, the
>>>>>>>>> cpsw_sl_ctl_set() function which writes the mac_control value performs a read
>>>>>>>>> modify write operation.
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Now, for SGMII, I would move setting CPSW_SL_CTL_EXT_EN to mac_config()
>>>>>>>>>> to enable in-band mode - don't we want in-band mode enabled all the
>>>>>>>>>> time while in SGMII mode so the PHY gets the response from the MAC?
>>>>>>>>>
>>>>>>>>> Thank you for pointing it out. I will move that to mac_config().
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Lastly, for RGMII at 10Mbps, you seem to suggest that you need RGMII
>>>>>>>>>> in-band mode enabled for that - but if you need RGMII in-band for
>>>>>>>>>> 10Mbps, wouldn't it make sense for the other speeds as well? If so,
>>>>>>>>>> wouldn't that mean that CPSW_SL_CTL_EXT_EN can always be set for
>>>>>>>>>> RGMII no matter what speed is being used?
>>>>>>>>>
>>>>>>>>> The CPSW MAC does not support forced mode at 10 Mbps RGMII. For this reason, if
>>>>>>>>> RGMII 10 Mbps is requested, it is set to in-band mode.
>>>>>>>>
>>>>>>>> What I'm saying is that if we have in-band signalling that is reliable
>>>>>>>> for a particular interface mode, why not always use it, rather than
>>>>>>>> singling out one specific speed as an exception? Does it not work in
>>>>>>>> 100Mbps and 1Gbps?
>>>>>>
>>>>>> While the CPSW MAC supports RGMII in-band status operation, the link partner
>>>>>> might not support it. I have also observed that forced mode is preferred to
>>>>>> in-band mode as implemented for another driver:
>>>>>> commit ade64eb5be9768e40c90ecb01295416abb2ddbac
>>>>>> net: dsa: microchip: Disable RGMII in-band status on KSZ9893
>>>>>>
>>>>>> and in the mail thread at:
>>>>>> https://lore.kernel.org/netdev/20200905160647.GJ3164319@xxxxxxx/
>>>>>> based on Andrew's suggestion, using forced mode appears to be better.
>>>>>>
>>>>>> Additionally, I have verified that switching to in-band status causes a
>>>>>> regression. Thus, I will prefer keeping it in forced mode for 100 and 1000 Mbps
>>>>>> RGMII mode which is the existing implementation in the driver. Please let me know.
>>>>>
>>>>> Okay, so what this seems to mean is if you have a PHY that does not
>>>>> support in-band status in RGMII mode, then 10Mbps isn't possible -
>>>>> because the MAC requires in-band status mode to select 10Mbps.
>>>>> To put it another way, in such a combination, 10Mbps link modes
>>>>> should not be advertised, nor should they be reported to userspace
>>>>> as being supported.
>>>>>
>>>>> Is that correct?
>>>>
>>>> Yes, if the PHY does not support in-band status, 10 Mbps RGMII will not work,
>>>> despite the MAC supporting 10 Mbps in-band RGMII. However, I notice the following:
>>>> If the RGMII interface speed is set to 10 Mbps via ethtool, but the:
>>>> managed = "in-band-status";
>>>> property is not mentioned in the device-tree, the interface is able to work with
>>>> 10 Mbps mode with the PHY. This is with the CPSW MAC configured for in-band mode
>>>> of operation at 10 Mbps RGMII mode. Please let me know what this indicates,
>>>> since it appears to me that 10 Mbps is functional in this special case (It might
>>>> be an erroneous configuration).
>>>
>>> I think you need to check carefully what is going on.
>>>
>>> Firstly, if you as the MAC is choosing to enable in-band status mode,
>>> but phylink isn't using in-band status mode, that is entirely a matter
>>> for your MAC driver.
>>>
>>> Secondly, you need to research what the PHY does during the inter-frame
>>> time (when in-band status would be transferred). This is when RX_CTL
>>> is 0,0, RX_DV is 0, RX_ER is 0.
>>>
>>> For in-band 10Mbps mode to work, RXD nibbles would need to be x001
>>> (middle two bits indicate RX clock = 2.5MHz clock for 10Mbps, lsb
>>> indicates link up). MSB determines duplex. Remember that 10Mbps can
>>> appear to work with mismatched duplex settings but can cause chaos on
>>> networks when it disagrees with what the rest of the network is doing.
>>>
>>> So, I think before one says "setting in-band mode for 10Mbps with a
>>> PHY that doesn't support in-band" really needs caution and research
>>> to check what _actually_ ends up happening, and whether it is really
>>> correct to do this.
>>
>> Thank you for the detailed explanation. I will analyze it and fix this. In the
>> meanwhile, is it acceptable for me to post the v2 of this series, with the other
>> suggestions implemented, while maintaining the status quo for the 10 Mbps RGMII
>> configuration in the driver? Please let me know.
> 
> Yes, but I would like a comment against the bit of code that enables
> in-band mode indicating that it's questionable whether it is correct.

Sure, thank you. I will add a TODO in that section, indicating that it needs to
be verified and fixed.

Regards,
Siddharth.