linux-kernel - Re: [PATCH] objtool: fix objtool regression on x32 systems

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Mon, 16 May 2022 23:25:17 +0200
From:   Peter Zijlstra <peterz@...radead.org>
To:     Mikulas Patocka <mpatocka@...hat.com>
Cc:     Linus Torvalds <torvalds@...ux-foundation.org>,
        Josh Poimboeuf <jpoimboe@...hat.com>,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH] objtool: fix objtool regression on x32 systems

On Mon, May 16, 2022 at 11:56:21AM -0400, Mikulas Patocka wrote:
> 
> 
> On Mon, 16 May 2022, Peter Zijlstra wrote:
> 
> > On Mon, May 16, 2022 at 11:06:36AM -0400, Mikulas Patocka wrote:
> > > The patch 4abff6d48dbc ("objtool: Fix code relocs vs weak symbols") makes
> > > the kernel unbootable.  The patch c087c6e7b551 ("objtool: Fix type of
> > > reloc::addend") attempts to fix it by replacing 'int' with 'long'.
> > > 
> > > However, we may be running on a system with x32 ABI and 'long' on x32 is
> > > 32-bit, thus the patch c087c6e7b551 doesn't really change anything and we
> > > still end up with miscompiled kernel.  This patch replaces 'long' with
> > > 'long long', so that the 64-bit kernel is correctly compiled on a x32
> > > system.
> > 
> > Hurmph.. you're building a 64bit kernel on a x32 hosted machine? And
> > this is the *only* thing that goes sideways?
> 
> I use the x32 ABI (that is 64-bit mode with addresses truncated to 
> 32-bit). gcc compiled for the x32 ABI is about 5% to 10% faster than gcc 
> compiled for the amd64 ABI. I installed the x32 gcc from Debian Ports.
> 
> > I suspect quite a bit of objtool assumes LP64 and won't quite work right
> > on ILP32 and we've just been lucky so far.
> 
> With this patch, the compiled kernel works. With kernels 5.17 or older, it 
> also works. I bisected it and the breakage is caused by the commit 
> 4abff6d48dbc.

Sure; but it works doesn't mean there aren't more latent issues. ILP32
hosted (cross) builds just aren't a thing I've ever considered. If we
really want to go support that then we should at least audit the whole
thing.

A quick look seems to suggest at least all the 'offset' fields should be
u64 or something. The only reason that works is because -mcmodel=kernel
keeps everything in the 2G range to make s32 immediates work. But it
isn't right.

Additionally, for things like LTO/IBT with vmlinux wide objtool runs,
i've seen objtool use up to ~20G of memory, I'm fairly sure x32 won't
cope with that.