[meta-virtualization] [PATCH V3 2/3] linux-yocto: make bbappend have effect conditionally

Fri Sep 29 20:12:30 PDT 2017

On 09/29/2017 04:01 AM, Bruce Ashfield wrote:
> On Wed, Sep 27, 2017 at 10:47 PM, ChenQi <Qi.Chen at windriver.com> wrote:
>> On 09/27/2017 08:43 PM, Bruce Ashfield wrote:
>>
>>
>>
>> On Tue, Sep 26, 2017 at 4:57 PM, Christopher Clark
>> <christopher.w.clark at gmail.com> wrote:
>>> On Mon, Sep 25, 2017 at 8:31 PM, Chen Qi <Qi.Chen at windriver.com> wrote:
>>>> Make these bbappend files to take effect only when DISTRO_FEATURES
>>>> contain 'virtualization'. Otherwise, we would meet failure failure
>>>> at system booting up qemux86. Related logs are as below.
>>>>
>>>>   systemd-modules-load[113]: Failed to insert 'kvm_amd': Operation not
>>>> supported
>>>>   systemd-modules-load[113]: Failed to insert 'kvm_intel': Operation not
>>>> supported
>>>>
>>>> Signed-off-by: Chen Qi <Qi.Chen at windriver.com>
>>> ...
>>>> diff --git a/recipes-kernel/linux/linux-yocto_virtualization.inc
>>>> b/recipes-kernel/linux/linux-yocto_virtualization.inc
>>>> new file mode 100644
>>>> index 0000000..8b296d7
>>>> --- /dev/null
>>>> +++ b/recipes-kernel/linux/linux-yocto_virtualization.inc
>>>> @@ -0,0 +1,20 @@
>>>> +FILESEXTRAPATHS_prepend := "${THISDIR}/linux-yocto:"
>>>> +
>>>> +SRC_URI += "file://xt-checksum.scc \
>>>> +            file://ebtables.scc \
>>>> +           file://vswitch.scc \
>>>> +           file://lxc.scc \
>>>> +            file://docker.scc \
>>>> +           "
>>>> +KERNEL_FEATURES_append = " features/kvm/qemu-kvm-enable.scc"
>>>> +
>>>> +KERNEL_MODULE_AUTOLOAD += "openvswitch"
>>>> +KERNEL_MODULE_AUTOLOAD += "kvm"
>>>> +KERNEL_MODULE_AUTOLOAD += "kvm-amd"
>>>> +KERNEL_MODULE_AUTOLOAD += "kvm-intel"
>>>> +
>>>> +# aufs kernel support required for xen-image-minimal
>>>> +KERNEL_FEATURES_append += "${@bb.utils.contains('DISTRO_FEATURES',
>>>> 'aufs', ' features/aufs/aufs-enable.scc', '', d)}"
>>>> +
>>>> +# xen kernel support
>>>> +SRC_URI += "${@bb.utils.contains('DISTRO_FEATURES', 'xen', '
>>>> file://xen.scc', '', d)}"
>>>> --
>>>> 2.11.0
>>>>
>>>> --
>>> Suggestion for a preferred change: make the KVM features depend on
>>> DISTRO_FEATURES containing 'kvm', in the same way that the Xen
>>> features do for DISTRO_FEATURES containing 'xen'.
>>>
>>> Here's why: those KVM kernel modules are not of interest on Xen
>>> systems. With this change, you're removing the KVM features from the
>>> qemu build to solve your bug in hand, by making it conditional on
>>> 'virtualization', but is still leaves them present when only Xen is
>>> needed.
>>>
>>> With that change, I'm not sure whether a 'virtualization' disto
>>> feature is immediately useful. It could make sense for features that
>>> are hypervisor-agnostic but I don't have an immediate example to point
>>> to.
>>
>> I like the suggestion as well. I'd like to future proof this a bit, and we
>> should keep the
>> virtualization feature around, but make the KVM features depend on kvm in
>> the
>> DISTRO_FEATURES. Less churn on distro flags is a good thing.
>>
>> Chen Qi, can you rework the check and warning message to reflect that 'kvm'
>> is the distro feature to set ?
>>
>> Bruce
>>
>>
>> Hi Bruce and Christopher,
>>
>> The status of 'kvm' in yocto (in specific, oe-core layer) is a little
>> different from that of 'xen'.
>>
>> 'xen' has been accepted as a distro feature, and the qemu recipe deals with
>> it like below.
>> meta/recipes-devtools/qemu/qemu.inc:
>> ${@bb.utils.filter('DISTRO_FEATURES', 'alsa xen', d)} \
>> meta/recipes-devtools/qemu/qemu.inc:PACKAGECONFIG[xen] =
>> "--enable-xen,--disable-xen,xen,xen-libxenstore xen-libxenctrl
>> xen-libxenguest"
>>
>> However, for 'kvm', things are different.
>> It's enabled by default for the qemu (native, nativesdk and target) like
>> below.
>> PACKAGECONFIG ??= " \
>>          fdt sdl \
>>          fdt sdl kvm \
>>          ${@bb.utils.filter('DISTRO_FEATURES', 'alsa xen', d)} \
>>          "
>> PACKAGECONFIG_class-native ??= "fdt alsa kvm"
>> PACKAGECONFIG_class-nativesdk ??= "fdt sdl kvm"
>> PACKAGECONFIG[kvm] = "--enable-kvm,--disable-kvm"
>> To make things more complicated, our runqemu script, and the related oeqa
>> parts also deal with the 'kvm' parameter.
>>
>> When talking about 'kvm' above, I'm talking about it from the view of
>> treating 'kvm' as a distro feature. I know that enabling kvm for qemu and
>> enabling kvm related configs for kernel are two different things, although
>> they are somewhat related.
>>
>> Given the above situation, I'd like to suggest we don't deal with 'kvm' in
>> meta-virtualization for now.
>> I'll find time to see if we can make 'kvm' as a distro feature in oe-core.
>> And we deal with 'kvm' in this layer after that.
> I'm actually not concerned about the oe-core part for now, getting this name
> into meta-virt correctly .. the first time .. is my concern. I don't
> want to only use
> 'virtualization' now, and then switch it later to match a new (more widely
> used) KVM distro feature.
>
> I had forgotten about the xen distro feature, and now that it has been pointed
> out, calling this 'virtualization' while it only impacts KVM build
> options to the
> kernel, doesn't make sense.
>
> So at a minimum, the kernel module autoload of kvm* should look for a kvm
> distro feature, but having the .inc included via the general 'virtualization'
> distro feature is still fine. Within the .inc, we have the xen and kvm
> differences
> .. which is fine.
>
> Bruce
>

Got it.
I had a sick leave yesterday.
I'll send out V4 soon.

Best Regards,
Chen Qi

>> Best Regards,
>> Chen Qi
>>
>>
>>
>>>
>>> Christopher
>>> --
>>> _______________________________________________
>>> meta-virtualization mailing list
>>> meta-virtualization at yoctoproject.org
>>> https://lists.yoctoproject.org/listinfo/meta-virtualization
>>
>>
>>
>> --
>> "Thou shalt not follow the NULL pointer, for chaos and madness await thee at
>> its end"
>>
>>
>
>