On Fri, May 7, 2021 at 2:42 AM Florian Weimer <fweimer(a)redhat.com> wrote:
We want to use kernel rebuilds as a gating test for toolchain
updates.
Unfortunately, per
Information for package kernel
<
https://koji.fedoraproject.org/koji/packageinfo?packageID=8>
the last successful rawhide kernel build was on 2021-04-28.
The question is what we should do about build failures like this.
Should we use non-rawhide kernels for our gating tests?
(The bug appears to be in BTF generation: it is not valid to assume that
static functions or variables are emitted under their declared names, or
that they have any particular calling convention or data layout. It's
probably best to drop the static if symbols are used for BTF extraction.
An alternative would be to use __attribute__ ((used)) instead, but then
the linker won't check for name collisions, which would result in
incorrect BTF.)
This long of an outage is rare, and the reasoning is because the fix is not
a kernel fix, it is in pahole. Jiri is aware, and has been for some time,
but upstream was trying to agree on which fix to use last I saw. Non
rawhide kernels are typically the safer bet, though there have been rare
cases where something needed to be added to the spec to make the kernel
build with a new toolchain (gcc 10), and that spec change was not
propagated down as that toolchain was never intended for the existing
stable releases. I think with the new ark workflow, it will not be a
problem.
Justin