[U-Boot] [PATCH] net/phy: Fix phy_connect() for phy addr 0

[Marek Vasut]

On 12/17/19 7:47 PM, Joe Hershberger wrote:
> On Tue, Dec 17, 2019 at 11:46 AM Marek Vasut <marex at denx.de> wrote:
>>
>> On 12/17/19 5:25 PM, Joe Hershberger wrote:
>>> Hi Marek,
>>
>> Hi Joe,
>>
>>> On Tue, Dec 17, 2019 at 1:39 AM Marek Vasut <marex at denx.de> wrote:
>>>>
>>>> On 11/7/19 9:04 PM, Joe Hershberger wrote:
>>>>> On Thu, Nov 7, 2019 at 1:16 PM Tom Rini <trini at konsulko.com> wrote:
>>>>>>
>>>>>> On Tue, Nov 05, 2019 at 04:05:11AM +0000, Priyanka Jain wrote:
>>>>>>
>>>>>>> Fix 'mask' calculation in phy_connect() for phy addr '0'.
>>>>>>> 'mask' is getting set to '0xffffffff' for phy addr '0'
>>>>>>> in phy_connect() whereas expected value is '0'.
>>>>>>>
>>>>>>>
>>>>>>> Signed-off-by: Priyanka Jain <priyanka.jain at nxp.com>
>>>>>>
>>>>>> Reported-by: tetsu-aoki via github
>>>>>
>>>>> Acked-by: Joe Hershberger <joe.hershberger at ni.com>
>>>>
>>>> Sadly, this breaks systems where a PHY is at address 0.
>>>> I have such an STM32MP1 system with LAN8720 PHY and since this patch, I
>>>> cannot use ethernet. Please revert.
>>>
>>> It seems like a case that shouldn't have worked before.
>>
>> Eh? PHY at address 0 definitely did work before and must work now.
> 
> Agreed that a phy at address 0 should work. Not agreed that because
> the value "0" used to work due to a bug that it must still. Which of
> these is the statement you are making? Do we already agree or
> disagree?

I am saying that because a board worked on rc4 and does not work on rc5,
this is a bug introduced by this patch in rc5 and must be fixed before
the release.

The address 0 is a PHY broadcast address for some PHYs, it's a fixed
address for other PHYs. Thus, a PHY at address 0 must work. If this is
broken now, it's a bug.

>>> What about
>>> this board requires the mask to be all 'f's, other than specifying the
>>> wrong phy address? It seems that in your case the phy address is not
>>> actually 0 (or the computed mask would find it), but your board dts
>>> may be setting it to 0 as an "unknown" value, but the correct unknown
>>> value should be "-1". It seems the issue is with these boards.
>>
>> Nope, the address is actually configured to 0 in hardware.
> 
> Can you double check that?

No, sorry, I know the hardware is fixed to 0. Checking it again will not
change this fact.

> The code as is should compute a mask of
> "0x01" which should match the offset for address 0. If it really is at
> 0 in hardware, maybe there is a different bug. Otherwise I don't see
> how this patch would work for the author.

Reverting this patch makes things work again for me.