[PATCH 1/5 v2] mv643xx_eth: add Device Tree bindings

[Florian Fainelli]

Hello Sebastian,

Le 04/05/13 15:58, Sebastian Hesselbarth a écrit :
> On Fri, Apr 5, 2013 at 11:56 AM, Florian Fainelli <florian at openwrt.org> wrote:
>> [snip]
>
> Florian,
>
> took me a while to try you patches out on Dove but now I fixed all
> issues. I will
> comment on all related patches but first I want to comment here.
>
> One general note for Dove related patches: You didn't remove the registration of
> ge platform_device from mach-dove/board-dt.c. That will lead to double
> registration
> of mdio and mv643xx_eth/shared, so you'll never be sure if DT or non-DT code is
> executed. I haven't checked mach-kirkwood/board-dt.c or orion5x code.

This was intentional, this patchset is just preparatory in the sense 
that it does no conversion of the existing users of the mv643xx_eth 
platform driver over DT (have some patches to that though). I wanted to 
resume the discussion on these bindings first, then proceed with the 
conversion.

>
>>>>          if (!mv643xx_eth_version_printed++)
>>>>                  pr_notice("MV-643xx 10/100/1000 ethernet driver version
>>>> %s\n",
>>>
>>>
>>> This is not related to your change, but there is a problem in this
>>> function that has already been discussed in the past if I remember
>>> correctly: The respective clock needs to be enabled here (at least
>>> on Kirkwood), since accesses to the hardware are done below.
>>> Enabling the clock only in mv643xx_eth_probe() is too late.
>>>
>>> As said, this is not a problem introduced by your changes (and which
>>> is currently circumvented by enabling the respective clocks in
>>> kirkwood_legacy_clk_init() and kirkwood_ge0x_init()), but we might
>>> want to fix this now to get rid of unconditionally enabling the GE
>>> clocks in the DT case.
>>
>>
>> I think there may have been some confusion between the "ethernet-group"
>> clock and the actual Ethernet port inside the "ethernet-group". The
>> mv643xx_eth driver assumes we have a per-port clock gating scheme, while I
>> think we have a per "ethernet-group" clock gating scheme instead. Like you
>> said, I think this should be addressed separately.
>
> IMHO, there should be a clocks property where ever you try to access registers,
> i.e. in all three "parts" mv643xx_eth_shared (group), mv643xx_eth
> (port) and mdio.
> Since port depends on shared it would be ok to have it per group but that may
> collide with other SoCs than Dove/Kirkwood that have per port clocks.

Ok, which means that we should also teach mv643xx_eth_shared_probe() 
about it, as well as the orion-mdio driver. I don't have any particular 
objections since it should just make things safer with respect to clocking.

>
> Is that separation (group/port) really required for any SoC?

Probably not, it was not clear when I looked at mv78xx0 if it uses two 
ports per group or 4 groups and 1 port. Anyway, since we are re-using 
the existing Device Tree binding definition and that the hardware 
present itself as ethernet groups and ports, I don't see any problem 
with keeping that difference since it allows for fine-grained 
representation of the hardware.

>
>>
>> [snip]
>>>
>>> You don't change the clk initialization here:
>>>
>>> #if defined(CONFIG_HAVE_CLK)
>>>          mp->clk = clk_get(&pdev->dev, (pdev->id ? "1" : "0"));
>>>          if (!IS_ERR(mp->clk)) {
>>>                  clk_prepare_enable(mp->clk);
>>>                  mp->t_clk = clk_get_rate(mp->clk);
>>>          }
>>> #endif
>>>
>>> Which, if I understand correctly, works in the DT case because you
>>> assign "clock-names" to the clocks in the DTS. However, I wonder
>>> whether this works for any but the first Ethernet device.
>
> Yes, it does. Assigned clocks from clocks property get a clock alias for
> that device name (node name). Using anything else than NULL here is
> IMHO just wrong. We should rather provide proper clock aliases for non-DT case.
>
>>> In the old platform device setup, the pdev->id was set when
>>> initialiazing the platform_device structure in common.c.  Where is
>>> this done in the DT case?
>>
>> Looks like you are right, in the DT case, I assume that we should lookup the
>> clock using NULL instead of "1" or "0" so we match any clock instead of a
>> specific one.
>
> Yes.
>
>> [snip]
>>>
>>>
>>> In phy_scan(), the phy is searched like this:
>>>
>>>                  snprintf(phy_id, sizeof(phy_id), PHY_ID_FMT,
>>>                                  "orion-mdio-mii", addr);
>>>
>>>                  phydev = phy_connect(mp->dev, phy_id,
>>> mv643xx_eth_adjust_link,
>>>                                  PHY_INTERFACE_MODE_GMII);
>>>
>>> But "orion-mdio-mii:xx" is the name of the PHY if MDIO is setup via a
>>> platform_device. I could not get this to work if the MDIO device is
>>> setup via DT. Am I doing something wrong?
>>
>> I just missed updating this part of the code to probe for PHYs. The board I
>> tested with uses a "PHY_NONE" configuration. I will add the missing bits for
>> of_phy_connect() to be called here.
>
> I don't think that the ethernet controller should probe the PHY's on mdio-bus
> at all. At least not for DT enabled platforms. I had a look at DT and non-DT
> mdio-bus sources, and realized that there is a bus scan for non-DT only.
> of_mdiobus_register requires you to set (and know) the PHY address.

One reason the Ethernet controller could do the probing is in the case 
we need to apply quirks (e.g: using phydev->flags) for instance. This 
can be done even after the MDIO bus driver did probe PHY devices though.

>
> I prepared a patch for of_mdio_register that will allow you to probe mdio and
> assign phy addresses to each node found. Currently, the heuristic for probing
> is: assign each phy node the next probed phy_addr starting with 0. But that
> will not allow to e.g. set some PHY addresses and probe the rest.

Ok, we just need to make sure that this does not break any specific use 
case, I don't think it does, since it seems to be more accurate or 
equivalent to Ethernet driver doing the probing.

>
> We had a similar discussion whether to probe or not for DT nodes, and I guess
> there also will be some discussion about the above patch. OTOH we could just
> (again) ask users of every kirkwood/orion5x/dove board to tell their
> phy addresses
> and fail to probe the phy for new boards...
>
> I will prepare a proper patch soon and post it on the corresponding lists.

Cool, thanks!

>
>>> Additionally, in phy_scan() there is this:
>>>
>>>          if (phy_addr == MV643XX_ETH_PHY_ADDR_DEFAULT) {
>>>                  start = phy_addr_get(mp) & 0x1f;
>>>                  num = 32;
>>>          } else {
>>>          ...
>>>
>>> MV643XX_ETH_PHY_ADDR_DEFAULT is defined as 0. However, many Kirkwood
>>> devices use "MV643XX_ETH_PHY_ADDR(0)".  If the module probe is
>>> deferred in mv643xx_eth because the MDIO driver is not yet loaded,
>>> all 32 PHY addresses are scanned without success.  This is not needed
>>> and clutters the log.
>>
>>
>> Ok, I am not sure how we can circumvent the log cluttering that happens,
>> what would be your suggestion?
>
> My suggestion is to change MV643XX_ETH_PHY_ADDR_DEFAULT from a valid
> phy address (0)
> to something invalid (32). I understand that using 0 helps if you
> don't want to set it in mv643xx's platform_data
> but it is always difficult to rely on that if 0 is a valid number.
>
> Changing the above to 32 should just work because most (all?) boards
> using phy_scan should also
> already use MV643XX_ETH_PHY_ADDR_DEFAULT. I also suggest to rename
> current define to a
> better name, e.g. MV643XX_ETH_PHY_ADDR_AUTOSCAN.

Sounds good to me.
--
Florian