Compile error ppc64le: Cannot find symbol for section 11: .text.unlikely.

[Coiby Xu]

On Fri, Feb 25, 2022 at 11:46:41AM +0800, Coiby Xu wrote:
>On Fri, Dec 03, 2021 at 04:54:19PM +0100, Veronika Kabatova wrote:
>>On Wed, Dec 1, 2021 at 3:20 AM Coiby Xu <coxu at redhat.com> wrote:
>>>
>>>On Wed, Nov 24, 2021 at 09:47:43PM +0800, Baoquan He wrote:
>>>>On 11/24/21 at 01:47pm, Veronika Kabatova wrote:
>>>>> Hi,
>>>>>
>>>>> for a while we've been seen the following error when compiling
>>>>> the mainline kernel with gcc 11.2 and binutils 2.37:
>>>>>
>>>>> 00:02:32 Cannot find symbol for section 11: .text.unlikely.
>>>>> 00:02:32 kernel/kexec_file.o: failed
>>>>> 00:02:32 make[3]: *** [scripts/Makefile.build:287: kernel/kexec_file.o] Error 1
>>>>> 00:02:32 make[3]: *** Deleting file 'kernel/kexec_file.o'
>>>>> 00:02:32 make[2]: *** [Makefile:1846: kernel] Error 2
>>>>> 00:02:32 make[2]: *** Waiting for unfinished jobs....
>>>>>
>>>>> The error only happens with ppc64le. I've tested this with cross
>>>>> compilation, but the only reference to the error I found suggests
>>>>> the same happens with the native compiles as well:
>>>>>
>>>>> https://github.com/groeck/linux-build-test/commit/142cbefbc0d37962c9a6c7f28ee415ecd5fd1e98
>>>>>
>>>>> In case it matters, the config used is the Fedora config with
>>>>> kselftest options enabled, which you can grab from
>>>>>
>>>>> https://gitlab.com/redhat/red-hat-ci-tools/kernel/cki-internal-pipelines/cki-trusted-contributors/-/jobs/1760752896/artifacts/raw/artifacts/kernel-mainline.kernel.org-ppc64le-e4e737bb5c170df6135a127739a9e6148ee3da82.config
>>>>>
>>>>>
>>>>> I've reached out to the Fedora compiler folks and Nick Clifton
>>>>> suggested this is a problem with the kernel:
>>>>>
>>>>>     This message comes from the recordmcount tool, which is part of the kernel
>>>>>     sources:
>>>>>
>>>>>     linux/scripts/recordmcount.[ch]
>>>>>
>>>>>     It appears to be triggered when a compiler update causes code to be
>>>>>     rearranged. The problem has been reported before in various forums,
>>>>>     but in particular I found this reference:
>>>>>
>>>>>     https://lore.kernel.org/lkml/20201204165742.3815221-2-arnd@kernel.org/
>>>>>
>>>>>     The point of which to me at least is that this is a kernel issue rather than
>>>>>     a compiler issue.  Ie there must be some weak symbols in kexec_file.o file
>>>>>     which need to be moved elsewhere.
>>>>
>>>>It could be arch_kexec_kernel_verify_sig() in kernel/kexec_file.c which
>>>>is __weak, but not implemented in any ARCH. If true, this has been
>>>>pointed out by Eric in one patch thread from Coiby.
>>>>
>>>>[PATCH v3 1/3] kexec: clean up arch_kexec_kernel_verify_sig
>>>>http://lkml.kernel.org/r/20211018083137.338757-2-coxu@redhat.com
>>>>
>>>>Maybe Coiby can fetch above config file and run the test to check.
>>>
>>>"[PATCH v3 1/3] kexec: clean up arch_kexec_kernel_verify_sig" alone
>>>would fix the error. If I turn arch_kexec_apply_relocations{_add,} into
>
>Sorry I meant "alone won't fix the error".
>
>>>static function, the error would be gone. As attached is the patch would
>>>make this error disappear.
>>>
>>
>>Thank you! I can confirm the attached patch fixes the problem.
>>
>>
>>Veronika
>>
>>>However, s390 and x86 have its own implementation of
>>>arch_kexec_apply_relocations_add. This makes it looks like to be gcc's
>>>issue.
>
>Based on the above point and further investigation, I think the root cause is
>find_secsym_ndx in linux/scripts/recordmcount.h,
> /*
>  * Find a symbol in the given section, to be used as the base for relocating
>  * the table of offsets of calls to mcount.  A local or global symbol suffices,
>  * but avoid a Weak symbol because it may be overridden; the change in value
>  * would invalidate the relocations of the offsets of the calls to mcount.
>  * Often the found symbol will be the unnamed local symbol generated by
>  * GNU 'as' for the start of each section.  For example:
>  *    Num:    Value  Size Type    Bind   Vis      Ndx Name
>  *      2: 00000000     0 SECTION LOCAL  DEFAULT    1
>  */
> static int find_secsym_ndx(unsigned const txtndx,
> 				char const *const txtname,
> 				uint_t *const recvalp,
> 				unsigned int *sym_index,
> 				Elf_Shdr const *const symhdr,
> 				Elf32_Word const *symtab,
> 				Elf32_Word const *symtab_shndx,
> 				Elf_Ehdr const *const ehdr)
> {
>        ...
>		if (txtndx == get_symindex(symp, symtab, symtab_shndx)
>			/* avoid STB_WEAK */
>
>	 fprintf(stderr, "Cannot find symbol for section %u: %s.\n",
>		txtndx, txtname);
>
>This function prints the above warning after failing to find
>arch_kexec_kernel_verify_sig or arch_kexec_apply_relocations{_add,} in
>section 11: .text.unlikely. because it ignores the weak symbol and ppc64le
>doesn't its arch implementations of these functions. I'll see if I can fix
>it in linux/scripts/recordmcount.h.

After digging deeper into linux/scripts/recordmcount.h, I think this
issue can be either fixed in the compiler or recordmcount. So I fild two bugs
- gcc: https://bugzilla.redhat.com/show_bug.cgi?id=2059838
- linux/scripts/recordmcount.h: https://bugzilla.redhat.com/show_bug.cgi?id=2059842

>
>>>
>>>
>>>>
>>>>Thanks
>>>>Baoquan
>>>>
>>>
>>>--
>>>Best regards,
>>>Coiby
>>
>
>-- 
>Best regards,
>Coiby

-- 
Best regards,
Coiby