Date:   Mon, 26 Oct 2020 16:45:53 -0700
From:   Kevin Hilman <khilman@...libre.com>
To:     Marc Zyngier <maz@...nel.org>
Cc:     Neil Armstrong <narmstrong@...libre.com>, jbrunet@...libre.com,
        linux-amlogic@...ts.infradead.org,
        linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 0/2] irq-meson-gpio: make it possible to build as a module

Marc Zyngier <maz@...nel.org> writes:

> On 2020-10-26 16:18, Kevin Hilman wrote:
>> Marc Zyngier <maz@...nel.org> writes:
>> 
>>> On Tue, 20 Oct 2020 08:25:30 +0100,
>>> Neil Armstrong <narmstrong@...libre.com> 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.

Yeah, I understand there's potential breakage things if you enable build
as a module, but that's not the default with this series so one would
have to do that on purpose.

If you don't manually enable as a module and leave the default as
built-in (e.g. CONFIG_MESON_IRQ_GPIO=y) does it work for you?  Using
your .config, I used the default to built-in and this series booted fine
for me on top of v5.10-rc1 on my VIM3L.

Using your .config, I do see issues with this driver when explicity
enabling as a module, and I'll debug those, but again, leaving it as
built-in, it's working fine for me with your .config.

I understand that even building as a built-in might change link order,
but I'd be pretty surprised if just changing from OF_DECLARE to a
built-in module made things break.

If so, I can commit to helping find/fix any issues there.

For that matter, I'm actively working on making all amlogic drivers
modular (including clocks, pinctrl, reset, etc.) and I have it all
working locally several SoCs, so I will also commit to actively help
find/fix any issues this uncovers with ordering.

My goal is to enable the *option* of building everything as modules (but
leave them as built-in by default) so I can help debug any remaining
fw_devlink issues.

Kevin

Powered by blists - more mailing lists