Re: [PATCH 0/2] irq-meson-gpio: make it possible to build as a module

[Marc Zyngier]

On 2020-10-26 16:18, Kevin Hilman wrote:
> Marc Zyngier <maz@xxxxxxxxxx> writes:
>
> > On Tue, 20 Oct 2020 08:25:30 +0100,
> > Neil Armstrong <narmstrong@xxxxxxxxxxxx> wrote:
> > > In order to reduce the kernel Image size on multi-platform 
> > > distributions,
> > > make it possible to build the Amlogic GPIO IRQ controller as a module
> > > by switching it to a platform driver.
> > >
> > > The second patch removes MESON_IRQ_GPIO selection from ARCH_MESON to 
> > > allow
> > > building the driver as module.
> > >
> > > Neil Armstrong (2):
> > >   irqchip: irq-meson-gpio: make it possible to build as a module
> > >   arm64: meson: remove MESON_IRQ_GPIO selection
> > >
> > >  arch/arm64/Kconfig.platforms     |  1 -
> > >  drivers/irqchip/Kconfig          |  5 +-
> > >  drivers/irqchip/irq-meson-gpio.c | 89 
> > > ++++++++++++++++++++------------
> > >  3 files changed, 59 insertions(+), 36 deletions(-)
> >
> > I've tried this series on my vim3l with the this driver compiled as a
> > module, and lost the Ethernet interface in the process, as the phy
> > wasn't able to resolve its interrupt and things fail later on:
> >
> > [   72.238291] meson8b-dwmac ff3f0000.ethernet eth1: no phy at addr -1
> > [   72.238917] meson8b-dwmac ff3f0000.ethernet eth1: stmmac_open: 
> > Cannot attach to PHY (error: -19)
> >
> > This is a generic problem with making DT-based interrupt controllers
> > modular when not *all* the drivers can deal with probing deferral.
>
> Yes, but this series still keeps the default as built-in.
>
> If you build as a module, and you add `fw_devlink=on` to the kernel
> command-line, device-links will be created based on DT dependencies
> which will ensure the right module load order.

It doesn't work here. I get the exact same error (well, with eth0 
instead
of eth1). In my experience, fw_devlink isn't reliable yet. Config on 
request.

> I've tested this series with `fw_devlink=on` on several Amlogic
> platforms and it works just fine, but since it requires the extra
> cmdline option, I think the default should remain built-in.
>
> So, I'd still like to see this series merged so that at least it's an
> option to enable this as a module.

I have taken similar patches in 5.9 for other SoC families (qcomm, mtk),
and ended up reverting them in -rc2, because there is simply too much
breakage. Even keeping it as built in changes the init order, which
tons of drivers depend on. I proposed a middle-of-the-road approach
(modules can break, built-in stays the same) which Rob pushed back on.

So either we fix fw_devlink to work for everything and be on by default,
or we keep the current setup.

> Also, another reason to make it optional is that not all platforms need
> this feature at all, but right now we select it for all Amlogic SoCs.

I understand that, but I don't want another episode of widespread
breakages, and this series definitely breaks things.

Thanks,

        M.
--
Jazz is not dead. It just smells funny...