[OE-core] [PATCH 0/3] Introduces kernel-initramfs recipe to resolve a implicit dependency issue

[Ming Liu]

On 01/19/2016 08:34 PM, Bruce Ashfield wrote:
> On 16-01-05 08:12 AM, Ming Liu wrote:
>> In current initramfs bundled kernel packaging policy, there are several
>> dependency chains co-existing:
>>
>> | "core-image-minimal.do_build" -> 
>> "core-image-minimal.do_bundle_initramfs"
>> | "core-image-minimal.do_bundle_initramfs" -> 
>> "virtual/kernel.do_bundle_initramfs"
>> | "core-image-minimal.do_bundle_initramfs" -> 
>> "core-image-minimal.do_rootfs"
>
> In current master, the above dependency should now be
> do_image_complete, correct ?
Yes, now it has been changed to do_image_complete.

>
>> | "core-image-minimal.do_rootfs" -> 
>> "virtual/kernel.do_package_write_${IMAGE_PKGTYPE}"
>
> And for the one above here, I'm not seeing this dependency. Are you
> saying that it comes via the do_image_complete dependency ?
It comes from the "recrdeptask" flag in image bbclass, for instance, 
meta/classes/rootfs_rpm.bbclass, it has:
do_rootfs[recrdeptask] += "do_package_write_rpm"

this makes sure that do_rootfs task of a certain image would run after 
do_package_write_rpm of all its DEPENDS and RDEPENDS, recursively.

>
>> | "virtual/kernel.do_package_write_${IMAGE_PKGTYPE}" -> 
>> "virtual/kernel.do_package"
>> | "virtual/kernel.do_package" -> "virtual/kernel.do_install"
>> |
>> | "virtual/kernel.do_deploy" -> "virtual/kernel.do_bundle_initramfs"
>> | "virtual/kernel.do_bundle_initramfs" -> "virtual/kernel.do_install"
>
> I'm somehow missing the above dependency as well. I suppose I could dump
> out the dot file, but I'd like to hear it explained here as well. Since
> if I can't get the dependency from the text of the commits, it will
> become hard to maintain.
They are intertask dependencies, I still take RPM as a example:
meta/classes/package_rpm.bbclass:addtask package_write_rpm after 
do_packagedata do_package
meta/classes/kernel.bbclass:addtask bundle_initramfs after do_install 
before do_deploy

>
>>
>> We could see, virtual/kernel.do_package is not explicitly depending on
>> virtual/kernel.do_bundle_initramfs so far, therefore, there is not a
>> workable way to add initramfs bundled kernel image into rootfs, because
>> kernel's do_bundle_initramfs could run parallelly with its do_package,
>> which means we will get a implicit kernel-image package that 
>> sometimes it
>> contains the initramfs bundled kernel or sometimes it doesn't.
>
> I do see what you are describing above. If we've defined
> INITRAMFS_IMAGE, the anonymous python in kernel.bbclass does make
> the kernel's do_bundle_initramfs depend on the initramfs image's
> do_image_complete.
>
> Why not just add a task dependency ?
Do you mean adding a task dependency between package and 
bundle_initramfs in kernel recipe? But that would introduce a circular 
dependency issue as described in commit log of 
609d5a9ab9e58bb1c2bcc2145399fbc8b701b85a, under following conditions:
1 Some kernel modules have been added into INITRAMFS_IMAGE.
2 INITRAMFS_IMAGE has been bundled into kernel image
3 kernel-image package has been added into IMAGE_INSTALL.

>
>>
>> To fix this problem, the idea is to let the initramfs bundled kernel
>> packaging depend on do_bundle_initramfs, meanwhile, to avoid the circular
>> dependency issue that commit: 609d5a9ab9e58bb1c2bcc2145399fbc8b701b85a
>> [ kernel.bbclass, image.bbclass: Implement kernel INITRAMFS 
>> dependency and bundling ]
>
> But here's my issue. We know that the INITRAMFS image cannot contain,
> or depend on a kernel itself. So the dependency won't be circular.
> It is true that we don't enforce that, but that has always been stated
> in the commits that created the bundling.
>
> Is it that condition you are trying to enforce with the split ?
I know that the users are not allowed to add kernel itself into a 
INITRAMFS_IMAGE meanwhile bundle it into kernel, which will certainly 
create a circular dependency. The split was not trying to resolve that 
condition, but to fix a implicit kernel-image package without 
introducing any circular dependencies.

//Ming Liu
>
> Bruce
>
>>
>> was trying to address, this dependency has to be splitted from kernel
>> recipe(at least, I could not figure out another way to achieve it), so a
>> new kernel-initramfs is introduced, in which a dependency chain is 
>> created:
>>
>> | "kernel-initramfs.do_install" -> "virtual/kernel.do_deploy"
>> | "virtual/kernel.do_deploy" -> "virtual/kernel.do_bundle_initramfs"
>>
>> Then the users can add initramfs bundled kernel image into rootfs by:
>>
>> IMAGE_INSTALL_append = " kernel-initramfs"
>>
>> without introducing any circular dependencies.
>>
>> Ming Liu (3):
>>    kernel.bbclass: do not install initramfs bundled kernel image
>>    image.bbclass: removes bundle_initramfs related code
>>    kernel-initramfs: new recipe, creates initramfs bundled kernel
>>      packaging
>>
>>   meta/classes/image.bbclass                    | 11 -----
>>   meta/classes/kernel.bbclass                   |  4 --
>>   meta/recipes-kernel/linux/kernel-initramfs.bb | 69 
>> +++++++++++++++++++++++++++
>>   3 files changed, 69 insertions(+), 15 deletions(-)
>>   create mode 100644 meta/recipes-kernel/linux/kernel-initramfs.bb
>>
>