Re: linux-next: Tree for Jul 31 - s390 crypto build breakage

[Ard Biesheuvel]

On Fri, 2 Aug 2019 at 03:20, Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx> wrote:
>
> Hi Herbert,
>
> On Thu, 1 Aug 2019 20:28:56 +0300 Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
> >
> > On Thu, 1 Aug 2019 at 15:28, Heiko Carstens <heiko.carstens@xxxxxxxxxx> wrote:
> > >
> > > On Wed, Jul 31, 2019 at 01:44:54PM +0200, Heiko Carstens wrote:
> > > > On Wed, Jul 31, 2019 at 09:32:16PM +1000, Herbert Xu wrote:
> > > > > On Wed, Jul 31, 2019 at 01:15:20PM +0200, Heiko Carstens wrote:
> > > > > >
> > > > > > However that doesn't fix the simd.h header file breakage with the
> > > > > > second patch :)
> > > > >
> > > > > That fix should be there now too.
> > > >
> > > > Yes, works now. Thank you!
> > >
> > > Still not... with linux-next as of today I get this (s390 defconfig):
> > >
> > > ERROR: "crypto_aegis128_decrypt_chunk_simd" [crypto/aegis128.ko] undefined!
> > > ERROR: "crypto_aegis128_update_simd" [crypto/aegis128.ko] undefined!
> > > ERROR: "crypto_aegis128_encrypt_chunk_simd" [crypto/aegis128.ko] undefined!
> > > scripts/Makefile.modpost:105: recipe for target 'modules-modpost' failed
> > >
> >
> > Hello Heiko,
> >
> > Apologies for the breakage. The first two fixes addressed obvious
> > shortcomings in my code, but with this issue, I'm a bit puzzled tbh.
> > The calls to these missing functions should be optimized away, since
> > have_simd never gets assigned if CONFIG_CRYPTO_AEGIS128_SIMD is not
> > defined, but for some reason, this isn't working. Which version of GCC
> > are you using?
> >
> > Also, could you please try whether the patch below fixes the problem? Thanks
> >
> > https://lore.kernel.org/linux-crypto/20190729074434.21064-1-ard.biesheuvel@xxxxxxxxxx/
>
> It might be time to revert all this series and try again.  The
> implementation seems to have not been well thought through from a kernel
> building point of view.  For a start the two commits
>
>   7cdc0ddbf74a ("crypto: aegis128 - add support for SIMD acceleration")
>   ecc8bc81f2fb ("crypto: aegis128 - provide a SIMD implementation based on NEON intrinsics")
>
> seem to be in the wrong order (function used in the first before being
> defined in the second).  There are a series of declarations of external
> functions in crypto/aegis128-core.c that should be in a header file.
> And there was the assumption that asm/simd.h was available everywhere.
>
> Also crypto_aegis128_decrypt_chunk_simd() is referenced in a structure
> initialisation (unprotected by any CONFIG_ variable - and so will be
> referenced even if it does not exist).  The compiler will have a hard
> time knowing that "have_simd" is effectively a constant zero (and
> crypto_simd_usable() is not constant).

The only assignment to have_simd is guarded by a if
(IS_ENABLED(CONFIG_xxx)) conditional, which is optimized away if the
Kconfig symbol is not set. Usually, the compiler uses this information
to infer that have_simd is a compile time constant '0', and optimizes
away all the code that depends on have_simd being true. I haven't
figured out yet why this doesn't work as expected on some versions of
GCC, since it is a very common pattern throughout the kernel.