Re: [lunn:v6.3-rc2-net-next-phy-leds 5/15] ld.lld: error: undefined symbol: devm_mdiobus_alloc_size

["Arnd Bergmann" <arnd@arndb.de>]

On Thu, Apr 20, 2023, at 15:07, Andrew Lunn wrote:
> On Thu, Apr 20, 2023 at 08:31:17AM +0200, Arnd Bergmann wrote:
>> 
>> It looks like this has hit linux-next now, I'm seeing the same problem in
>> my own randconfig builds after Andrew's 01e5b728e9e4 ("net: phy: Add a
>> binding for PHY LEDs").
>
> Hi Arnd
>
> I tried to fix this with:
>
> commit 37f9b2a6c086bb28487a0682b8098f907861c4a1
> Author: Andrew Lunn <andrew@lunn.ch>
> Date:   Thu Mar 30 20:13:29 2023 -0500
>
>     net: ethernet: Add missing depends on MDIO_DEVRES
>    
>     A number of MDIO drivers make use of devm_mdiobus_alloc_size(). This
>     is only available when CONFIG_MDIO_DEVRES is enabled. Add missing
>     depends or selects, depending on if there are circular dependencies or
>     not. This avoids linker errors, especially for randconfig builds.
>
> All the failures i've seen have CONFIG_MDIO_DEVRES set to disabled. So
> i added either depends on, or select to the drivers which use it. I
> missed the LAN743x.

I've sent a different patch now. From what I can tell, your patch does
not address the actual problem, it just works around one of the symptoms:

When you have a driver that selects PHYLIB, but PHYLIB itself
depends on something that is not enabled, Kconfig just gives up
and stops handling any other 'select' statements from that chain.

selecting MDIO_DEVRES from each of the drivers works around this
one thing, but it does not solve the actual dependency, for that you
would have to add 'depends on LED_CLASS' on each of the network
drivers as well, which is clearly not the intention.

>> The problem here is that both PHYLIB and LEDS_CLASS are user-visible
>> tristate symbols that are referenced from other Kconfig symbols with
>> both 'depends on' and 'select'. Having the two interact introduces a
>> number of ways that lead to circular dependencies.
>
> I was getting circular dependencies with first versions of the above
> patch. I initially tried depends on everywhere, and then had to drop
> back to select for a few cases.
>
>> It might be ok to use 'select LEDS_CLASS' from PHYLIB, but I have not
>> tried that yet and I expect this will result in other build failures.
>> 
>> A better solution would be to change all drivers that currently use
>> 'select PHYLIB' to 'depends on PHYLIB' and have PHYLIB itself
>> 'default ETHERNET' to avoid most of the regressions, but doing this
>> for 6.4 is a bit risky and can cause other problems.
>
> For 6.4, will adding more depend on/select MDIO_DEVRES help? That
> should be low risk.

I think the best way is to drop your MDIO_DEVRES patch and instead
apply mine (or some variation of that) from

https://lore.kernel.org/lkml/20230420084624.3005701-1-arnd@kernel.org/

Once any missing or recursive dependencies are handled, the devres
problem should be fixed as well. I have completed around 150
randconfig builds with that patch and have not seen any further
problems.

     Arnd