[oe] [RFC] initramfs support for linux-handhelds-2.6

[Paul Sokolovsky]

Hello,

On Mon, 25 Feb 2008 10:15:11 +0100
Koen Kooi <koen at dominion.kabel.utwente.nl> wrote:

> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> Paul Sokolovsky schreef:
> | Hello,
> |
> | Failing to add generic and fully automated kernel-with-initramfs
> build, | I have patch to add external initramfs to kernel for
> linux-hh. I'd | still like to establish conventions for other kernel
> recipes to follow, | so here're the patch for suggestion. The idea is
> simple: |
> | 1. KERNEL_INITRAMFS_PATH is set to the full path of cpio.gz image to
> | use as initramfs. Empty value (default) means don't use initramfs.
> If | file doesn't exist, build fails with an error message (content
> is left | to specific recipe - there unfortunately doesn't seem to be
> a way to | sensibly provide detailed one in generic case).
> |
> | 2. KERNEL_INITRAMFS_PATH is considered a distro realm, like for
> example | kernel PREFERRED_VERSION. Below, angstrom distro config
> sets it.
> 
> | +KERNEL_INITRAMFS_PATH =
> "${TMPDIR}/deploy/uclibc/images/${MACHINE}/initramfs-bootmenu-image-${MACHINE}.cpio.gz"
> 
> | +	if [ -n "${KERNEL_INITRAMFS_PATH}" -a "${ANGSTROM_MODE}"
> == "glibc" ]; then
> 
> A number of things:
> 
> 1) Make the initramfs generation a seperate function (or task) that
> can be called at will, instead of always being called in
> do_compile_prepend 

Initramfs *generation* is a separate task. With this scheme, it is
completely external, binary blob drop. Except by setting initramfs path
to point to OE's deploy dir, one can use previous OE build result. So,
for our case, it would be:

ANGSTROM_MODE=uclibc IMAGE_FSTYPES="cpio.gz" bitbake \
	initramfs-bootmenu-image
bitbake x11-image

and everything will magically work.


> 2) Keep it simple and keep the kernel and
> initramfs in the same libc realm, so a uclibc based initramfs needs a
> uclibc kernel (I known there is no such thing, but you get the idea).

Keeping it separate is exactly *the simple* thing. Trying to do it
otherwise calls for hard-to-resolve-in-good-way complications. Again,
there's recursion in the job of building kernel with initramfs:

To build kernel, you need to have initramfs. To build initramfs, you
need modules. To build modules, you need to build kernel.

You saw one solution for breaking this vicious cycle - the recursion
was lashed into OE's own blackbox where it was kept under control under
given, real-world achievable circumstances. That was rejected.

Current solution splits cycle the other way: 1) "uclibc" kernel is never
built with initramfs; 2) that means that its build finishes without
recursion, giving out modules; 3) uclibc initramfs is built with that
modules; 4) then "glibc" kernel recipes pulls it automatically(!) from
uclibc deploy dir, builds kernel with it, and more importantly
*packages* it w/o any dirty magic.

> That way you can just fish it out of DEPLOY_DIR_IMAGE

The main problem is not kernel image, but kernel package.

> 3) Add an initramfs-<kernel-name_version>.bb that does
> 
> ~  require <kernel-name_version>.bb
> ~  FILESPATH = "/path/magic/whatever"
> ~  do_compile_prepend() { do_initramfs}
> 
> This involves some extra build time ('uclibc' kernel build and a
> 'uclibc' initramfs kernel build), but is more transparent and less
> error prone than your plan. I think it even avoids the dependency
> problems :)

Packaging. 

Besides, it's conceptually not very correct. For a given
configuration (distro+machine), there used to be one kernel, right?
Now, there're two, one of which is not supposed to be used, nor it's
even working (now that we rely on initramfs, a kernel w/o it is dead
beef). One way to ensure transparency is to not confuse people with
giving out build results which are not intended to be there. My patch
works exactly that way - if Angstrom currently really supports glibc as
"production" config, and thus "glibc" kernel is production one, then it
will be always built with initramfs, and non-initramfs (== broken)
kernel won't be built at all, to not make all this prone to errors like
it getting into rootfs, picked up by users, etc.

Yes, that's not too general. But as I told, the correct way to solve
this is to have zImage vs modules split for kernel recipe. (Currently
tried splits - user level vs OE balckbox; uclibc vs glibc - are just
hacks around the issue).

> 
> regards,
> 
> Koen
> 
[]

-- 
Best regards,
 Paul                          mailto:pmiscml at gmail.com