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

On 10 January 2017 at 19:22, Andy Lutomirski [off-list ref] wrote:

On Tue, Jan 10, 2017 at 11:16 AM, Ard Biesheuvel
[off-list ref] wrote:

quoted

On 10 January 2017 at 19:00, Andy Lutomirski [off-list ref] wrote:

quoted

On Tue, Jan 10, 2017 at 9:30 AM, Ard Biesheuvel
[off-list ref] wrote:

quoted

On 10 January 2017 at 14:33, Herbert Xu [off-list ref] wrote:

quoted

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.

Actually, the breakage is introduced by the patch Herbert refers to

https://patchwork.kernel.org/patch/9468391/

where the state is replaced by a simple

u32 state[16] __aligned(CHACHA20_STATE_ALIGN);

which seemed harmless enough to me. So the code above works fine.

So how about just the one-line patch of adding the
force_align_arg_pointer?  Would that solve the problem?

If it does what it says on the tin, it should fix the issue, but after
adding the attribute, I get the exact same object output, so there's
something dodgy going on here.