Re: Re: linux-next: build warning after merge of the bpf-next tree

[Palmer Dabbelt]

On Fri, 10 Jan 2020 14:28:17 PST (-0800), alexandre@xxxxxxxx wrote:
> Hi guys,
>
> On 10/27/19 8:02 PM, Stephen Rothwell wrote:
> > Hi all,
> >
> > On Fri, 18 Oct 2019 10:56:57 +1100 Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx> wrote:
> > > Hi all,
> > >
> > > After merging the bpf-next tree, today's linux-next build (powerpc
> > > ppc64_defconfig) produced this warning:
> > >
> > > WARNING: 2 bad relocations
> > > c000000001998a48 R_PPC64_ADDR64    _binary__btf_vmlinux_bin_start
> > > c000000001998a50 R_PPC64_ADDR64    _binary__btf_vmlinux_bin_end
> > >
> > > Introduced by commit
> > >
> > >    8580ac9404f6 ("bpf: Process in-kernel BTF")
> > This warning now appears in the net-next tree build.
> I bump that thread up because Zong also noticed that 2 new relocations for
> those symbols appeared in my riscv relocatable kernel branch following
> that commit.
>
> I also noticed 2 new relocations R_AARCH64_ABS64 appearing in arm64 kernel.
>
> Those 2 weak undefined symbols have existed since commit
> 341dfcf8d78e ("btf: expose BTF info through sysfs") but this is the fact
> to declare those symbols into btf.c that produced those relocations.
>
> I'm not sure what this all means, but this is not something I expected
> for riscv for
> a kernel linked with -shared/-fpie. Maybe should we just leave them to
> zero ?
>
> I think that deserves a deeper look if someone understands all this
> better than I do.

Can you give me a pointer to your tree and how to build a relocatable kernel?
Weak undefined symbols have the absolute value 0, but the kernel is linked at
an address such that 0 can't be reached by normal means.  When I added support
to binutils for this I did it in a way that required almost no code --
essetially I just stopped dissallowing x0 as a possible base register for PCREL
relocations, which results in 0 always being accessible.  I just wanted to get
the kernel to build again, so I didn't worry about chasing around all the
addressing modes.  The PIC/PIE support generates different relocations and I
wouldn't be surprised if I just missed one (or more likely all) of them.

It's probably a simple fix, though I feel like every time I say that about the
linker I end up spending a month in there...