2019-08-26 16:37:02

by He Zhe

[permalink] [raw]
Subject: kernel/bpf/core.o: warning: objtool: ___bpf_prog_run.cold()+0x7: call without frame pointer save/setup

Hi All,

Since 3193c0836f20 ("bpf: Disable GCC -fgcse optimization for ___bpf_prog_run()"),
We have got the following warning,
kernel/bpf/core.o: warning: objtool: ___bpf_prog_run.cold()+0x7: call without frame pointer save/setup

If reverting the above commit, we will get the following warning,
kernel/bpf/core.o: warning: objtool: ___bpf_prog_run()+0x8b9: sibling call from callable instruction with modified stack frame
if CONFIG_RETPOLINE=n, and no warning if CONFIG_RETPOLINE=y


Zhe


2019-08-27 15:00:20

by Josh Poimboeuf

[permalink] [raw]
Subject: Re: kernel/bpf/core.o: warning: objtool: ___bpf_prog_run.cold()+0x7: call without frame pointer save/setup

On Tue, Aug 27, 2019 at 10:43:27AM +0800, He Zhe wrote:
>
>
> On 8/26/19 11:18 PM, Josh Poimboeuf wrote:
> > On Mon, Aug 26, 2019 at 10:42:53PM +0800, He Zhe wrote:
> >> Hi All,
> >>
> >> Since 3193c0836f20 ("bpf: Disable GCC -fgcse optimization for ___bpf_prog_run()"),
> >> We have got the following warning,
> >> kernel/bpf/core.o: warning: objtool: ___bpf_prog_run.cold()+0x7: call without frame pointer save/setup
> >>
> >> If reverting the above commit, we will get the following warning,
> >> kernel/bpf/core.o: warning: objtool: ___bpf_prog_run()+0x8b9: sibling call from callable instruction with modified stack frame
> >> if CONFIG_RETPOLINE=n, and no warning if CONFIG_RETPOLINE=y
> > Can you please share the following:
> >
> > - core.o file
>
> Attached.
>
> >
> > The following would also be helpful for me to try to recreate it:
> >
> > - config file
> > - compiler version
> > - kernel version
>
> I pasted them in the other reply.

Thanks. I was able to recreate. I reduced it to:

void a(b);
__attribute__((optimize(""))) c(void) { a(); }

Apparently '__attribute__((optimize()))' is overwriting GCC cmdline
flags, including -fno-omit-frame-pointer. I had assumed it would append
instead of replace.

I'm guessing this is a GCC "feature" instead of a bug. I'll need to
follow up.

--
Josh