On Tue, 11 Mar 2025 at 15:12, Venkat Rao Bagalkote <venkat88@xxxxxxxxxxxxx> wrote: > > > On 10/03/25 2:15 pm, Alexei Starovoitov wrote: > > On Mon, Mar 10, 2025 at 8:32 AM Venkat Rao Bagalkote > > <venkat88@xxxxxxxxxxxxx> wrote: > >> Greetings!!! > >> > >> selftests/bpf fails to compile with below error on bpf-next repo with > >> commit head: f28214603dc6c09b3b5e67b1ebd5ca83ad943ce3 > >> > >> Repo link: > >> https://web.git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/ > >> > >> Reverting below commit resolves the issue. > >> > >> Commit ID: 48b3be8d7f82bea6affe6b9f11ee67380b55ede8 > > ... > > > >> If you happen to fix the issue, please add below tag. > >> > >> Reported-by: Venkat Rao Bagalkote <venkat88@xxxxxxxxxxxxx> > > Not quite. The issue is likely that your llvm is too old. > > Please upgrade. > > > Thanks for the feedback. > > I did try the compilation on fedora41, with linux-mainline kernel and > /sefltests/bpf compiled successfully. But on the same set-up > /selftests/bpf failed to compile on bpf-next kernel. > > > OS: Fedora Linux 41 (Server Edition) > LLVM: llvm-19.1.7-3.fc41.ppc64le After staring at the error log in disbelief, I guess I have the answer. https://elixir.bootlin.com/linux/v6.13.6/source/arch/powerpc/include/asm/qspinlock_types.h#L8 PowerPC overrides the type such that lock->val is no longer atomic_t. All of the bpf_atomic.h helpers assume it's atomic_t, hence they try to do lock->val.counter which obviously fails with val is u32 instead of a struct. The "fix" would probably be to replace the usage of struct qspinlock with our own definition copying the asm-generic qspinlock type. I can send a patch later today in the evening, unless you beat me to it. Thanks > > gcc version 14.2.1 20250110 (Red Hat 14.2.1-7) (GCC) > > Passing repo: > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git > Failing repo: > https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/ > > Regards, > > Venkat. > > >
