Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Daniel Axtens <dja@axtens.net>
To:     Jan Stancek <jstancek@redhat.com>, mpe@ellerman.id.au,
        benh@kernel.crashing.org, paulus@samba.org,
        christophe.leroy@csgroup.eu, linuxppc-dev@lists.ozlabs.org
Cc:     linux-kernel@vger.kernel.org, jstancek@redhat.com
Subject: Re: [PATCH/RFC] powerpc/module_64: allow .init_array constructors
 to run
In-Reply-To: <920acea9aa18e4f2956581a8e158bdaa376fdf63.1629203945.git.jstancek@redhat.com>
References: <920acea9aa18e4f2956581a8e158bdaa376fdf63.1629203945.git.jstancek@redhat.com>
Date:   Wed, 18 Aug 2021 13:32:43 +1000
Message-ID: <877dgjbf6s.fsf@dja-thinkpad.axtens.net>
MIME-Version: 1.0
Content-Type: text/plain
Precedence: bulk

Jan Stancek <jstancek@redhat.com> writes:

> gcov and kasan rely on compiler generated constructor code.
> For modules, gcc-8 with gcov enabled generates .init_array section,
> but on ppc64le it doesn't get executed. find_module_sections() never
> finds .init_array section, because module_frob_arch_sections() renames
> it to _init_array.
>
> Avoid renaming .init_array section, so do_mod_ctors() can use it.

This (the existing renaming) breaks a KASAN test too, so I'd love to see
this fixed.

However, I don't know that renaming the section is a complete fix: from
memory it is still be possible to get relocations that are too far away
and require separate trampolines. But I wasn't able to construct a
module that exhibited this behaviour and test a fix before I got pulled
away to other things.

Kind regards,
Daniel

>
> Cc: Michael Ellerman <mpe@ellerman.id.au>
> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> Cc: Paul Mackerras <paulus@samba.org>
> Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
> Signed-off-by: Jan Stancek <jstancek@redhat.com>
> ---
> I wasn't able to trace the comment:
>   "We don't handle .init for the moment: rename to _init"
> to original patch (it pre-dates .git). I'm not sure if it
> still applies today, so I limited patch to .init_array. This
> fixes gcov for modules for me on ppc64le 5.14.0-rc6.
>
> Renaming issue is also mentioned in kasan patches here:
>   https://patchwork.ozlabs.org/project/linuxppc-dev/cover/20210319144058.772525-1-dja@axtens
>
>  arch/powerpc/kernel/module_64.c | 10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
>
> diff --git a/arch/powerpc/kernel/module_64.c b/arch/powerpc/kernel/module_64.c
> index 6baa676e7cb6..c604b13ea6bf 100644
> --- a/arch/powerpc/kernel/module_64.c
> +++ b/arch/powerpc/kernel/module_64.c
> @@ -299,8 +299,16 @@ int module_frob_arch_sections(Elf64_Ehdr *hdr,
>  					  sechdrs[i].sh_size);
>  
>  		/* We don't handle .init for the moment: rename to _init */
> -		while ((p = strstr(secstrings + sechdrs[i].sh_name, ".init")))
> +		while ((p = strstr(secstrings + sechdrs[i].sh_name, ".init"))) {
> +#ifdef CONFIG_CONSTRUCTORS
> +			/* find_module_sections() needs .init_array intact */
> +			if (strstr(secstrings + sechdrs[i].sh_name,
> +				".init_array")) {
> +				break;
> +			}
> +#endif
>  			p[0] = '_';
> +		}
>  
>  		if (sechdrs[i].sh_type == SHT_SYMTAB)
>  			dedotify((void *)hdr + sechdrs[i].sh_offset,
> -- 
> 2.27.0