[OE-core] [v2 PATCH] kernel.bbclass, image.bbclass: Implement kernel INITRAMFS dependency and bundling

Fri Sep 27 09:05:28 UTC 2013

On Fri, 2013-09-27 at 10:07 +0200, Andrea Adami wrote:
> On Sat, Aug 24, 2013 at 7:15 PM, Richard Purdie
> <richard.purdie at linuxfoundation.org> wrote:
> > On Thu, 2013-08-22 at 18:04 -0500, Jason Wessel wrote:
> >> This patch aims to fix the following two cases for the INITRAMFS generation.
> >>   1) Allow an image recipe to specify a paired INITRAMFS recipe such
> >>      as core-image-minimal-initramfs.  This allows building a base
> >>      image which always generates the needed initramfs image in one step
> >>   2) Allow building a single binary which contains a kernel and
> >>      the initramfs.
> >>
> >> A key requirement of the initramfs is to be able to add kernel
> >> modules.  The current implementation of the INITRAMFS_IMAGE variable
> >> has a circular dependency when using kernel modules in the initramfs
> >> image.bb file that is caused by kernel.bbclass trying to build the
> >> initramfs before the kernel's do_install rule.
> >>
> >> The solution for this problem is to have the kernel's
> >> do_bundle_initramfs_image task depend on the do_rootfs from the
> >> INITRAMFS_IMAGE and not some intermediate point.  The image.bbclass
> >> will also sets up dependencies to make the initramfs creation task run
> >> last.
> >>
> >> The code to bundle the kernel and initramfs together has been added.
> >> At a high level, all it is doing is invoking a second compilation of
> >> the kernel but changing the value of CONFIG_INITRAMFS_SOURCE to point
> >> to the generated initramfs from the image recipe.
> >>
> >> [YOCTO #4072]
> >>
> >> Signed-off-by: Jason Wessel <jason.wessel at windriver.com>
> >> Acked-by: Bruce Ashfield <bruce.ashfield at windriver.com>
> >
> > I have a couple of things I'd really like to see get resolved here. One
> > is below, the other is I'm worried about the packaged output differences
> > since we can package the kernel into a package file and now its going to
> > be different.
> >
> > I appreciate its a hard problem to solve but not impossible. Basically
> > we move the package generation for that single package into a separate
> > recipe and have it depend on the bundling task if/as/when needed. The
> > bundle task stashes the kernel in the sysroot, the other recipe simply
> > packages it. Its a little bit of a dance but should ensure we get
> > everything consistent.
> >
> >
> >> ---
> >>  meta/classes/image.bbclass  |   12 ++++++
> >>  meta/classes/kernel.bbclass |   96 +++++++++++++++++++++++++++++++++++++------
> >>  meta/conf/local.conf.sample |   20 +++++++++
> >>  3 files changed, 116 insertions(+), 12 deletions(-)
> >>
> >> diff --git a/meta/classes/image.bbclass b/meta/classes/image.bbclass
> >> index 909702a..23967ec 100644
> >> --- a/meta/classes/image.bbclass
> >> +++ b/meta/classes/image.bbclass
> >> @@ -130,6 +130,10 @@ python () {
> >>      d.setVar('MULTILIB_VENDORS', ml_vendor_list)
> >>
> >>      check_image_features(d)
> >> +    initramfs_image = d.getVar('INITRAMFS_IMAGE', True) or ""
> >> +    if initramfs_image != "":
> >> +        d.appendVarFlag('do_build', 'depends', " %s:do_bundle_initramfs" %  d.getVar('PN', True))
> >> +        d.appendVarFlag('do_bundle_initramfs', 'depends', " %s:do_rootfs" % initramfs_image)
> >>  }
> >>
> >>  #
> >> @@ -629,3 +633,11 @@ do_package_write_deb[noexec] = "1"
> >>  do_package_write_rpm[noexec] = "1"
> >>
> >>  addtask rootfs before do_build
> >> +# Allow the kernel to be repacked with the initramfs and boot image file as a single file
> >> +do_bundle_initramfs[depends] += "virtual/kernel:do_bundle_initramfs"
> >> +do_bundle_initramfs[nostamp] = "1"
> >> +do_bundle_initramfs[noexec] = "1"
> >> +do_bundle_initramfs () {
> >> +     :
> >> +}
> >> +addtask bundle_initramfs after do_rootfs
> >> diff --git a/meta/classes/kernel.bbclass b/meta/classes/kernel.bbclass
> >> index e039dfc..8cf66ce 100644
> >> --- a/meta/classes/kernel.bbclass
> >> +++ b/meta/classes/kernel.bbclass
> >> @@ -9,6 +9,7 @@ INHIBIT_DEFAULT_DEPS = "1"
> >>  KERNEL_IMAGETYPE ?= "zImage"
> >>  INITRAMFS_IMAGE ?= ""
> >>  INITRAMFS_TASK ?= ""
> >> +INITRAMFS_IMAGE_BUNDLE ?= ""
> >>
> >>  python __anonymous () {
> >>      kerneltype = d.getVar('KERNEL_IMAGETYPE', True) or ''
> >> @@ -19,7 +20,15 @@ python __anonymous () {
> >>
> >>      image = d.getVar('INITRAMFS_IMAGE', True)
> >>      if image:
> >> -        d.setVar('INITRAMFS_TASK', '${INITRAMFS_IMAGE}:do_rootfs')
> >> +        d.appendVarFlag('do_bundle_initramfs', 'depends', ' ${INITRAMFS_IMAGE}:do_rootfs')
> >> +
> >> +    # NOTE: setting INITRAMFS_TASK is for backward compatibility
> >> +    #       The preferred method is to set INITRAMFS_IMAGE, because
> >> +    #       this INITRAMFS_TASK has circular dependency problems
> >> +    #       if the initramfs requires kernel modules
> >> +    image_task = d.getVar('INITRAMFS_TASK', True)
> >> +    if image_task:
> >> +        d.appendVarFlag('do_configure', 'depends', ' ${INITRAMFS_TASK}')
> >>  }
> >>
> >>  inherit kernel-arch deploy
> >> @@ -72,9 +81,82 @@ KERNEL_SRC_PATH = "/usr/src/kernel"
> >>
> >>  KERNEL_IMAGETYPE_FOR_MAKE = "${@(lambda s: s[:-3] if s[-3:] == ".gz" else s)(d.getVar('KERNEL_IMAGETYPE', True))}"
> >>
> >> +copy_initramfs() {
> >> +     echo "Copying initramfs into ./usr ..."
> >> +     # Find and use the first initramfs image archive type we find
> >> +     rm -f ${B}/usr/${INITRAMFS_IMAGE}-${MACHINE}.cpio
> >> +     for img in cpio.gz cpio.lzo cpio.lzma cpio.xz; do
> >> +             if [ -e "${DEPLOY_DIR_IMAGE}/${INITRAMFS_IMAGE}-${MACHINE}.$img" ]; then
> >> +                     cp ${DEPLOY_DIR_IMAGE}/${INITRAMFS_IMAGE}-${MACHINE}.$img ${B}/usr/.
> >> +                     case $img in
> >> +                     *gz)
> >> +                             echo "gzip decompressing image"
> >> +                             gunzip -f ${B}/usr/${INITRAMFS_IMAGE}-${MACHINE}.$img
> >> +                             break
> >> +                             ;;
> >> +                     *lzo)
> >> +                             echo "lzo decompressing image"
> >> +                             lzop -df ${B}/usr/${INITRAMFS_IMAGE}-${MACHINE}.$img
> >> +                             break
> >> +                             ;;
> >> +                     *lzma)
> >> +                             echo "lzma decompressing image"
> >> +                             lzmash -df ${B}/usr/${INITRAMFS_IMAGE}-${MACHINE}.$img
> >> +                             break
> >> +                             ;;
> >> +                     *xz)
> >> +                             echo "xz decompressing image"
> >> +                             xz -df ${B}/usr/${INITRAMFS_IMAGE}-${MACHINE}.$img
> >> +                             break
> >> +                             ;;
> >> +                     esac
> >> +             fi
> >> +     done
> >> +     echo "Finished copy of initramfs into ./usr"
> >> +}
> >
> > But what about my bzip2'd image? ;-)
> >
> > I'd suggest we rid of this and instead ensure that we're generating an
> > uncompressed cpio image. The image generation code will happily sort
> > that our for us if we ask it for that specific image type.
> >
> > I'd also wondered if we could remove INITRAMFS_TASK since its just going
> > to confuse things and I'd prefer to maintain only one way of doing this
> > if at all possible.
> >
> > Cheers,
> >
> > Richard
> 
> 
> Though our observations, in an effort to solve Yocto #4072 the
> patchset has been merged.
> As side effect it broke the 'old style' creation of initramfs so some
> recipes are now unbuildable.
> 
> I'm in touch with Bruce and Jason and I have already a patch for
> kernel.bbclass restoring the old functionality by adding one more
> variable in each recipe: INITRAMFS_TASK.
> 
> I'm pretty sure we could have built the same kind of images containing
> modules in one pass before as well and still dislike the idea of
> having to set a variable in local.conf to spread it to the kernel and
> to the image (to any image...)
> 
> With 1.5 approaching I'd like to have the issue solved as soon as possible.
> Richard, when is deadline for core-patches?

Basically ASAP. The next release build is the start of next week but any
patches going into that need to have been run through an autobuilder
cycle first.

This is the first I've heard of the problem, other than some comments
about possible problems on irc. Can someone please open up a bug for
this and clearly describe what used to work and now doesn't and what the
possible fixes or workarounds are?

I don't think anyone likes regressions however if we don't have the open
bug, its very hard to track.

You say you've talked to Bruce/Jason but why didn't the discussion
happen on the mailing list? That way others may have been able to help
and now I could read back the list and see for myself what the issue is.
As it is, I simply don't know other than the need to set a new variable
which is hinted at above...

Cheers,

Richard