Re: [PATCH] arm64/sve: fix genksyms generation

[Will Deacon]

On Mon, Jun 17, 2019 at 02:21:46PM +0200, Arnd Bergmann wrote:
> On Mon, Jun 17, 2019 at 1:26 PM Will Deacon <will.deacon@xxxxxxx> wrote:
> > On Mon, Jun 17, 2019 at 12:42:11PM +0200, Arnd Bergmann wrote:
> > > genksyms does not understand __uint128_t, so we get a build failure
> > > in the fpsimd module when it cannot export a symbol right:
> >
> > The fpsimd code is builtin, so which module is actually failing? My
> > allmodconfig build succeeds, so I must be missing something.
> 
> It happened for me on randconfig builds, you can find one such configuration
> at https://pastebin.com/cU8iQ4ta now. I was building this with clang
> rather than gcc, which may affect the issue, but I assumed not.

Hmm, I've failed to reproduce the issue with that config and either GCC
(7.1.1 and 8.3.0) or Clang (a flavour of 9.0.0 from a few months ago).

> > > WARNING: EXPORT symbol "kernel_neon_begin" [vmlinux] version generation failed, symbol will not be versioned.
> > > /home/arnd/cross/x86_64/gcc-8.1.0-nolibc/aarch64-linux/bin/aarch64-linux-ld: arch/arm64/kernel/fpsimd.o: relocation R_AARCH64_ABS32 against `__crc_kernel_neon_begin' can not be used when making a shared object
> > > arch/arm64/kernel/fpsimd.o:(.data+0x0): dangerous relocation: unsupported relocation
> > > arch/arm64/kernel/fpsimd.o:(".discard.addressable"+0x0): dangerous relocation: unsupported relocation
> > > arch/arm64/kernel/fpsimd.o:(".discard.addressable"+0x8): dangerous relocation: unsupported relocation
> > >
> > > We could teach genksyms about the type, but it's easier to just
> > > work around it by defining that type locally in a way that genksyms
> > > understands.
> > >
> > > Fixes: 41040cf7c5f0 ("arm64/sve: Fix missing SVE/FPSIMD endianness conversions")
> >
> > I can't see which part of that patch causes the problem, so I'm a bit wary
> > of the fix. We've been using __uint128_t for a while now, and I see there's
> > one in the x86 kvm code as well, so it would be nice to understand what's
> > happening here so that we can avoid running into it in future as well.
> 
> The problem is only in files that export a symbol. This is also the
> case in arch/x86/kernel/kvm.c, but it may be lucky because the
> type only appears /after/ the last export in that file.
> 
> > > Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> > > ---
> > >  arch/arm64/kernel/fpsimd.c | 3 +++
> > >  1 file changed, 3 insertions(+)
> > >
> > > diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
> > > index 07f238ef47ae..2aba07cccf50 100644
> > > --- a/arch/arm64/kernel/fpsimd.c
> > > +++ b/arch/arm64/kernel/fpsimd.c
> > > @@ -400,6 +400,9 @@ static int __init sve_sysctl_init(void) { return 0; }
> > >  #define ZREG(sve_state, vq, n) ((char *)(sve_state) +                \
> > >       (SVE_SIG_ZREG_OFFSET(vq, n) - SVE_SIG_REGS_OFFSET))
> > >
> > > +#ifdef __GENKSYMS__
> > > +typedef __u64 __uint128_t[2];
> > > +#endif
> >
> > I suspect I need to figure out what genksyms is doing, but I'm nervous
> > about exposing this as an array type without understanding whether or
> > not that has consequences for its operation.
> 
> The entire point is genksyms is to ensure that types of exported symbols
> are compatible. To do this, it has a limited parser for C source code that
> understands the basic types (char, int, long, _Bool, etc) and how to
> aggregate them into structs and function arguments. This process has
> always been fragile, and it clearly breaks when it fails to understand a
> particular type.

Ok, but the patch that appears to cause this problem doesn't change the
type of anything we're exporting. The symbol in your log is
"kernel_neon_begin" which is:

	void kernel_neon_begin(void);

so I'm still fairly confused about the problem. In fact, even if I create
a silly:

	void will_kernel_neon_begin(__uint128_t);

function, then somehow I see it being processed:

	__crc_will_kernel_neon_begin = 0x5401d250;

Is there some way that your passing '-w' to genksyms?

Will