Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 1 Jul 2022 14:23:39 -0700
From:   Luis Chamberlain <mcgrof@kernel.org>
To:     Helge Deller <deller@gmx.de>,
        Lucas De Marchi <lucas.demarchi@intel.com>
Cc:     jeyu@kernel.org, linux-modules@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-parisc@vger.kernel.org
Subject: Re: [PATCH v2] modules: Ensure natural alignment for
 .altinstructions and __bug_table sections
Message-ID: <Yr9l24rvCAPJvuJQ@bombadil.infradead.org>
References: <Yr8/gr8e8I7tVX4d@p100>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <Yr8/gr8e8I7tVX4d@p100>
Sender: Luis Chamberlain <mcgrof@infradead.org>
Precedence: bulk

On Fri, Jul 01, 2022 at 08:40:02PM +0200, Helge Deller wrote:
> In the kernel image vmlinux.lds.S linker scripts the .altinstructions
> and __bug_table sections are 32- or 64-bit aligned because they hold 32-
> and/or 64-bit values.
> 
> But for modules the module.lds.S linker script doesn't define a default
> alignment yet, so the linker chooses the default byte alignment, which
> then leads to unnecessary unaligned memory accesses at runtime.
> 
> Usually such unaligned accesses are unnoticed, because either the
> hardware (as on x86 CPUs) or in-kernel exception handlers (e.g. on hppa
> or sparc) emulate and fix them up at runtime.
> 
> On hppa the 32-bit unalignment exception handler was temporarily broken
> due another bad commit, and as such wrong values were returned on
> unaligned accesses to the altinstructions table.

OK so some bad commit broke something which caused bad alignment access
on altinstructions... But why on modules?!

I am not aware of modules using alternatives, given that alternatives
are hacks to help with bootup. For modules we can use other things
like jump labels, static keys.

So I don't understand still how this happened yet.

> This then led to
> undefined behaviour because wrong kernel addresses were patched and we
> suddenly faced lots of unrelated bugs, as can be seen in this mail
> thread:
> https://lore.kernel.org/all/07d91863-dacc-a503-aa2b-05c3b92a1e39@bell.net/T/#mab602dfa32be5e229d5e192ab012af196d04d75d
> 
> This patch adds the missing natural alignment for kernel modules to
> avoid unnecessary (hard- or software-based) fixups.

Is it correct to infer that issue you found through a bad commit was
then through code inspection after the bad commit made the kernel do
something stupid with unaligned access to some module altinstructions
section ? Ie, that should not have happened.

I'd like to determine if this is a stable fix, a regression, etc. And
this is not yet clear.

  Luis

> 
> Signed-off-by: Helge Deller <deller@gmx.de>
> ---
>  scripts/module.lds.S | 2 ++
>  1 file changed, 2 insertions(+)
> 
> --
> v2: updated commit message
> 
> diff --git a/scripts/module.lds.S b/scripts/module.lds.S
> index 1d0e1e4dc3d2..3a3aa2354ed8 100644
> --- a/scripts/module.lds.S
> +++ b/scripts/module.lds.S
> @@ -27,6 +27,8 @@ SECTIONS {
>  	.ctors			0 : ALIGN(8) { *(SORT(.ctors.*)) *(.ctors) }
>  	.init_array		0 : ALIGN(8) { *(SORT(.init_array.*)) *(.init_array) }
> 
> +	.altinstructions	0 : ALIGN(8) { KEEP(*(.altinstructions)) }
> +	__bug_table		0 : ALIGN(8) { KEEP(*(__bug_table)) }
>  	__jump_table		0 : ALIGN(8) { KEEP(*(__jump_table)) }
> 
>  	__patchable_function_entries : { *(__patchable_function_entries) }