Re: [RFC PATCH 00/23] arm: defconfigs: use kconfig fragments

From: Olof Johansson
Date: Wed Dec 07 2016 - 16:28:47 EST


On Wed, Dec 7, 2016 at 1:07 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> On Wednesday, December 7, 2016 12:41:29 PM CET Bartlomiej Zolnierkiewicz wrote:
>>
>> On Tuesday, December 06, 2016 11:03:34 AM Olof Johansson wrote:
>> > On Tue, Dec 6, 2016 at 4:38 AM, Bartlomiej Zolnierkiewicz
>> > <b.zolnierkie@xxxxxxxxxxx> wrote:
>> > > Hi,
>> > >
>> > > This RFC patchset starts convertion of ARM defconfigs to use kconfig
>> > > fragments and dynamically generate defconfigs. The goals of this
>> > > work are to:
>> >
>> > You don't provide any motivation as to why this is better. As far as I
>>
>> Benefits are:
>>
>> - no code duplication (this initial patchset alone removes ~1700 lines
>> from defconfigs without any change in functionality)
>
> This may be interesting

Management of the fragments is the big headache here. I haven't seen
any system that does it well downstream either in a way that scales as
far as we'd need it to.

>> - prevention of "multi" defconfigs (i.e. multi_v7_defconfig) going out
>> of sync with "SoC-family" ones (i.e. exynos_defconfig) - there will
>> be just one place to update when changing things
>
> I'm not convinced this is worthwhile: in a lot of cases, the soc-specific
> configs want to enable things built-in, while the more generic ones
> tend to use loadable modules.

Agreed.

>> - possibility to add support for more optimized defconfigs (i.e. board
>> specific ones) in the future without duplicating the code
>
> I'd prefer seeing fewer top-level options than more of them, so
> this doesn't really help.
>
>> - making it easier to define arch specific parts of defconfigs in
>> the future if we decide on doing it (i.e. we may want to enable
>> things like CONFIG_SYSVIPC for all defconfigs)
>
> The example you give is for something that should be decided
> in architecture-independent Kconfig language rather than
> per architecture, and that won't require fragments.
>
>> > am concerned it'll just be a mess.
>> >
>> > So:
>> >
>> > Nack. So much nack. I really don't want to see a proliferation of
>> > config fragments like this.
>> >
>> > I had a feeling it was a bad idea to pick up that one line config
>> > fragment before, since it opened the door for this kind of mess.
>>
>> Like I said in the cover-letter I'm not satisfied with the current
>> patches and they have much room for improvement.
>>
>> However I see that you don't like the idea itself...
>
> I do think that there is some room for more config fragments in
> mainline, but not most of the patches you have here. Some areas
> that I think would benefit from fragments are:
>
> - architecture level selection: v6/v6k/v7/v7ve/v8 could have a
> common defconfig file that starts out with all v6+ enabled,
> but then having fragments that disable the older architectures
> and platforms using them while turning on features that are only
> available on newer architectures
>
> - A "debug" fragment would be nice, to turn on common options that
> add a lot of useful runtime checks at the expense of performance
> or code size.

Hmm, some of these might work but several useful debug options (in
particular DEBUG_LL for early errors) are per-system/platform.

> - A "distro" fragment that turns on all loadable modules that are
> enabled by common distributions (e.g. two or more of
> debian/fedora/opensuse/gentoo), to let you build a drop-in
> replacement kernel for a shipping distro. This would also allow
> the distros to strip their own config files and just specify
> whatever differs from the others.

Keeping this in sync with the distro kernel could be a bit awkward
(and possibly churny).


-Olof