Re: [PATCH] recordmcount: Support empty section from recent binutils

Le 29/03/2022 à 00:31, Joel Stanley a écrit :

On Mon, 29 Nov 2021 at 22:43, Christophe Leroy
[off-list ref] wrote:

quoted

Le 29/11/2021 à 18:43, Steven Rostedt a écrit :

quoted

On Fri, 26 Nov 2021 08:43:23 +0000
LEROY Christophe [off-list ref] wrote:

quoted

Le 24/11/2021 à 15:43, Christophe Leroy a écrit :

quoted

Looks like recent binutils (2.36 and over ?) may empty some section,
leading to failure like:

     Cannot find symbol for section 11: .text.unlikely.
     kernel/kexec_file.o: failed
     make[1]: *** [scripts/Makefile.build:287: kernel/kexec_file.o] Error 1

In order to avoid that, ensure that the section has a content before
returning it's name in has_rel_mcount().

This patch doesn't work, on PPC32 I get the following message with this
patch applied:

[    0.000000] ftrace: No functions to be traced?

Without the patch I get:

[    0.000000] ftrace: allocating 22381 entries in 66 pages
[    0.000000] ftrace: allocated 66 pages with 2 groups

Because of this report, I have not applied this patch (even though I was
about to push it to Linus).

I'm pulling it from my queue until this gets resolved.

I have no idea on how to fix that for the moment.

With GCC 10 (binutils 2.36) an objdump -x on kernel/kexec_file.o gives:

0000000000000000 l    d  .text.unlikely 0000000000000000 .text.unlikely
0000000000000000  w    F .text.unlikely 0000000000000038
.arch_kexec_apply_relocations_add
0000000000000038  w    F .text.unlikely 0000000000000038
.arch_kexec_apply_relocations


With GCC 11 (binutils 2.37) the same gives:

0000000000000000  w    F .text.unlikely 0000000000000038
.arch_kexec_apply_relocations_add
0000000000000038  w    F .text.unlikely 0000000000000038
.arch_kexec_apply_relocations


The problem is that recordmcount drops weak symbols, and it doesn't find
any non-weak symbol in .text.unlikely

Explication given at
https://elixir.bootlin.com/linux/v5.16-rc2/source/scripts/recordmcount.h#L506

I have no idea on what to do.

Did you end up finding a solution for this issue?

Not really, my solution was to switch to the kernel compiler at 
https://mirrors.edge.kernel.org/pub/tools/crosstool/ which embeds 
binutils 2.36

But it looks like using objtool instead of recordmcount doesn't exhibit 
the problem.

https://patchwork.ozlabs.org/project/linuxppc-dev/patch/20220318105140.43914-4-sv@linux.ibm.com/

Christophe