Description of problem:

somehow, the version was 'missing' from the broken build.
Things get *really* unhappy when uname returns null.
Sometimes you see recursive segfaults. Sometimes you see "kernel too old"
messages, sometimes, you see nothing at all, just a hang.

Reviewing the older 686-smp kernels that had similar problems showed that they
too had broken utsname's.

Checking the build logs did show that there was indeed some differences in the
generation of version.h between the two builds.
It also only seemed to happen when the build system is *really* busy. Like say,
during a mass-rebuild of every package. Or when it's near a freeze, and everyone
wants to get their fixes into the buildsystem asap. Perfect timing for a bug of
this nature.

So what was the actual cause ?

The kernel specfile did a build in the following way..

make -s ARCH=$Arch nonint_oldconfig > /dev/null
make -s ARCH=$Arch include/linux/version.h
make -s ARCH=$Arch %{?_smp_mflags} %{make_target}
make -s ARCH=$Arch %{?_smp_mflags} modules || exit 1


The 2nd line was added at some point to work around some bug or other which
subsequently was fixed upstream. (Lesson #2: Document *why* you're adding a
workaround). It's totally unnecessary now. A majority of the time, it was
harmless. But under the right conditions, it did *totally* the wrong thing, and
you ended up with an empty .kernelrelease

Version-Release number of selected component (if applicable):
2.6.9-34.EL

How reproducible:

Boot a i686-smp kernel



Steps to Reproduce:
1.build a kernel
2.boot the kernel randomly.
3.
  
Actual results:

booting mysteriously panics

Expected results:

booting not mysteriously panic

Additional info: