linux-kernel - Re: x86-64: Maintain 16-byte stack alignment

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:   Tue, 10 Jan 2017 11:00:31 -0800
From:   Andy Lutomirski <luto@...capital.net>
To:     Ard Biesheuvel <ard.biesheuvel@...aro.org>
Cc:     Herbert Xu <herbert@...dor.apana.org.au>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Linux Crypto Mailing List <linux-crypto@...r.kernel.org>,
        Linus Torvalds <torvalds@...ux-foundation.org>,
        Ingo Molnar <mingo@...nel.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        Andy Lutomirski <luto@...nel.org>
Subject: Re: x86-64: Maintain 16-byte stack alignment

On Tue, Jan 10, 2017 at 9:30 AM, Ard Biesheuvel
<ard.biesheuvel@...aro.org> wrote:
> On 10 January 2017 at 14:33, Herbert Xu <herbert@...dor.apana.org.au> wrote:
>> I recently applied the patch
>>
>>         https://patchwork.kernel.org/patch/9468391/
>>
>> and ended up with a boot crash when it tried to run the x86 chacha20
>> code.  It turned out that the patch changed a manually aligned
>> stack buffer to one that is aligned by gcc.  What was happening was
>> that gcc can stack align to any value on x86-64 except 16.  The
>> reason is that gcc assumes that the stack is always 16-byte aligned,
>> which is not actually the case in the kernel.
>>
>
> Apologies for introducing this breakage. It seemed like an obvious and
> simple cleanup, so I didn't even bother to mention it in the commit
> log, but if the kernel does not guarantee 16 byte alignment, I guess
> we should revert to the old method. If SSE instructions are the only
> ones that require this alignment, then I suppose not having a ABI
> conforming stack pointer should not be an issue in general.

Here's what I think is really going on.  This is partially from
memory, so I could be off base.  The kernel is up against
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383, which means that,
on some GCC versions (like the bad one and maybe even current ones),
things compiled without -mno-sse can't have the stack alignment set
properly.  IMO we should fix this in the affected code, not the entry
code.  In fact, I think that fixing it in the entry code won't even
fully fix it because modern GCC will compile the rest of the kernel
with 8-byte alignment and the stack will get randomly unaligned (GCC
4.8 and newer).

Can we just add __attribute__((force_align_arg_pointer)) to the
affected functions?  Maybe have:

#define __USES_SSE __attribute__((force_align_arg_pointer))

on affected gcc versions?

***HOWEVER***

I think this is missing the tree for the supposed forest.  The actual
affected code appears to be:

static int chacha20_simd(struct blkcipher_desc *desc, struct scatterlist *dst,
                         struct scatterlist *src, unsigned int nbytes)
{
        u32 *state, state_buf[16 + (CHACHA20_STATE_ALIGN / sizeof(u32)) - 1];

...

        state = (u32 *)roundup((uintptr_t)state_buf, CHACHA20_STATE_ALIGN);

gcc presumably infers (incorrectly) that state_buf is 16-byte aligned
and optimizes out the roundup.  How about just declaring an actual
__aligned(16) buffer, marking the function
__attribute__((force_align_arg_pointer)), and being done with it?
After all, we need that forcible alignment on *all* gcc versions.

--Andy