[OE-core] distro/machine:kernel feature mapping and version checking

[Hart, Darren]

On Fri, 2013-11-22 at 11:05 +0000, Paul Eggleton wrote:
> Hi Darren,
> 
> On Thursday 21 November 2013 18:47:07 Hart, Darren wrote:
> > Regarding the following 2 bugs:
> > 
> > https://bugzilla.yoctoproject.org/show_bug.cgi?id=5574
> > Add kernel version / configuration check mechanism
> > 
> > and 
> > 
> > https://bugzilla.yoctoproject.org/show_bug.cgi?id=2267
> > Integrate DISTRO_FEATURES with KERNEL_FEATURES
> > 
> > There is a need to reduce errors where a DISTRO config might break due
> > to lack of kernel support, things such as systemd. There is also
> > interest in enabling certain kernel features based on
> > DISTRO/MACHINE_FEATURES, such as wifi.
> > 
> > Neither of these should depend on the Kernel Version as even with the
> > right version, if the CONFIG_* feature is missing, the image will not
> > work correctly. There is also the risk of false negatives when a feature
> > has been backported to a kernel version that didn't have the feature
> > previously.
> 
> Right; I hadn't appreciated this when I wrote the bug but we definitely don't 
> want to be checking versions.

Yup, just provided for context.

> 
> > The recommended approach would be to check for the required CONFIG_*
> > options after the linux-yocto configuration stage.
> > 
> > There is also the topic of DISTRO_FEATURES ~= "wifi" impacting how the
> > kernel will be built. The MACHINE should have some say in how this is
> > done - if the machine can never have wifi, building wifi into the kernel
> > doesn't make a lot of sense. Something like the following might make
> > sense:
> > 
> > for FEATURE in DISTRO_FEATURES:
> >     if MACHINE_FEATURES contains FEATURES:
> >         KERNEL_FEATURES += FEATURE_override
> > 
> > Where FEATURE_override is defined something like this:
> > 
> > FEATURE = FEATURE_default
> > if exists FEATURE_distro:
> >     FEATURE = FEATURE_distro
> > if exists FEATURE_machine:
> >     FEATURE = FEATURE_machine
> 
> There are a couple of slight quirks here:
> 
> 1) Not all features match up between DISTRO_FEATURES and MACHINE_FEATURES like 
> this, just a select list. (This is why we currently have COMBINED_FEATURES.)

systemd is a good example I guess, the MACHINE doesn't need to support
it. However for a DISTRO with wifi, it still doesn't make sense to
install all those packages or build the kernel drivers if the board has
no wifi and no expansion capability. Would an opt-out model be more
practical?

if FEATURE in DISTRO_FEATURES:
    if not MACHINE_DENY_FEATURES contains FEATURE:
        KERNEL_FEATURES += FEATURE_override

?

> 
> 2) There is some handling in packagegroup-base.bbclass to enable 
> wifi/bluetooth/3g/nfc functionality in userspace if these features are *not* in 
> MACHINE_FEATURES but they are in DISTRO_FEATURES *and* MACHINE_FEATURES 
> indicates some means of expansion exists e.g. pci, pcmcia, usbhost. Whether we 
> want to try to support this at the kernel level I don't know.

Hrm... this appears broken to me. If the MACHINE doesn't feature them,
then the kernel is unlikely to support them, so regardless of if the
expansion physically exists, the device still will not work on that
image. If a feature requiring hardware/driver support is not listed in
MACHINE_FEATURES, userspace infrastructure probably should not be added.
Generic BSPs should be adding these MACHINE_FEATURES, specific BSPs
don't care about being generic.... so... it works out I think.

> 
> > This is effectively a fragment name which needs to be provided by the
> > linux-yocto kernel meta data as it will be kernel version dependent.
> > 
> > The linkage I'm not sure about is how to know what to test for in the
> > kernel.bbclass without intimate knowledge of the kernel version CONFIG
> > options in recipe space.
> 
> So are you convinced you want to try to solve these two with one solution in 
> kernel.bbclass rather than giving individual recipes some easy methods of 
> checking kernel config options? 

Not at all. Looking to the experts such as yourself to help iterate to a
reasonable implementation.

We could do something like provide a kernel.bbclass interface to
checking for specific FEATURE options, which each kernel recipe
(version) could override... I think.


kernel_has_wireless()
  if ${B}/.config ~= "^CONFIG_WIRELESS=[my]"
  ...
      return True
  return False

(yeah.... babel)

This would add a fairly nasty dependency on virtual/kernel:configure for
any package that wanted to use those functions. This may be inevitable
unfortunately.

> I agree they are related, but I had imagined 
> that we'd solve them both and solve them separately; only doing the feature 
> map when we're building the kernel still leaves us with recipes like udev that 
> aren't controlled by a specific feature but do still have kernel config 
> requirements (of course, you could just assume that if you're using sysvinit 
> then you're using udev and handle the issue that way, but that seems like a 
> hack to me).
> 
> Cheers,
> Paul

OK, let's keep going, how do we get to solving them independently from
here?

Thanks Paul!

-- 
Darren Hart
Intel Open Source Technology Center
Yocto Project - Linux Kernel