Re: [PATCH 0/2] ftrace/recordmcount: Handle object files without section symbols

[Naveen N. Rao]

Hi Steven,

Steven Rostedt wrote:
> On Wed, 27 Apr 2022 15:01:20 +0530
> "Naveen N. Rao" <naveen.n.rao@xxxxxxxxxxxxxxxxxx> wrote:
>
> > This solves a build issue on powerpc with binutils v2.36 and newer [1].
> > Since commit d1bcae833b32f1 ("ELF: Don't generate unused section
> > symbols") [2], binutils started dropping section symbols that it thought
> > were unused. Due to this, in certain scenarios, recordmcount is unable 
> > to find a non-weak symbol to generate a relocation record against.
> >
> > Clang integrated assembler is also aggressive in dropping section 
> > symbols [3].
> >
> > In the past, there have been various workarounds to address this. See 
> > commits 55d5b7dd6451b5 ("initramfs: fix clang build failure") and 
> > 6e7b64b9dd6d96 ("elfcore: fix building with clang") and a recent patch:
> > https://lore.kernel.org/linuxppc-dev/20220425174128.11455-1-naveen.n.rao@xxxxxxxxxxxxxxxxxx/T/#u
> >
> > Fix this issue by using the weak symbol in the relocation record. This 
> > can result in duplicate locations in the mcount table if those weak 
> > functions are overridden, so have ftrace skip dupicate entries.
> >
> > Objtool already follows this approach, so patch 2 updates recordmcount 
> > to do the same. Patch 1 updates ftrace to skip duplicate entries.
> >
> > - Naveen
> >
> >
> > [1] https://github.com/linuxppc/issues/issues/388
> > [2] https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=d1bcae833b32f1
> > [3] https://github.com/ClangBuiltLinux/linux/issues/981
>
> There's been work to handle weak functions, but I'm not sure that work
> handled the issues here. Are these patches still needed, or was there
> another workaround to handle the problems this addressed?

I'm afraid these patches are still needed to address issues in 
recordmcount.

I submitted patches to remove use of weak functions in the kexec 
subsystem, but those have only enabled building ppc64le defconfig 
without errors:
https://lore.kernel.org/all/20220519091237.676736-1-naveen.n.rao@xxxxxxxxxxxxxxxxxx/
https://lore.kernel.org/all/cover.1656659357.git.naveen.n.rao@xxxxxxxxxxxxxxxxxx/

The patch adding support for FTRACE_MCOUNT_MAX_OFFSET to powerpc only 
helps ignore weak functions during runtime:
https://lore.kernel.org/all/20220809105425.424045-1-naveen.n.rao@xxxxxxxxxxxxxxxxxx/

We still see errors from recordmcount when trying to build certain 
powerpc configs.

We are pursuing support for objtool, which doesn't have the same issues:
https://lore.kernel.org/all/20220808114908.240813-1-sv@xxxxxxxxxxxxx/


- Naveen