Re: [PATCHv2 bpf-next 1/3] bpf, tests: tweak endianness selection

[Stanislav Fomichev <sdf@fomichev.me>]

On 03/21, Daniel Borkmann wrote:
> On 03/20/2019 11:45 PM, Yonghong Song wrote:
> > On 3/20/19 3:27 PM, Stanislav Fomichev wrote:
> >> On 03/20, Yonghong Song wrote:
> >>> On 3/20/19 10:13 AM, Stanislav Fomichev wrote:
> >>>> On 03/20, Sergey Senozhatsky wrote:
> >>>>> Not all compilers have __builtin_bswap16() and __builtin_bswap32(),
> >>>>> thus not all compilers are able to compile the following code:
> >>>>>
> >>>>>           (__builtin_constant_p(x) ? \
> >>>>>                   ___constant_swab16(x) : __builtin_bswap16(x))
> >>>>>
> >>>>> That's the reason why bpf_ntohl() doesn't work on GCC < 4.8, for
> >>>>> instance:
> >>>>>
> >>>>>           error: implicit declaration of function '__builtin_bswap16'
> >>>>>
> >>>>> We can use __builtin_bswap16() only if compiler has this built-in,
> >>>>> that is, only if __HAVE_BUILTIN_BSWAP16__ is defined. Standard UAPI
> >>>>> __swab16()/__swab32() take care of that, and, additionally, handle
> >>>>> __builtin_constant_p() cases as well:
> >>>>>
> >>>>>    #ifdef __HAVE_BUILTIN_BSWAP16__
> >>>>>    #define __swab16(x) (__u16)__builtin_bswap16((__u16)(x))
> >>>>>    #else
> >>>>>    #define __swab16(x)                             \
> >>>>>            (__builtin_constant_p((__u16)(x)) ?     \
> >>>>>            ___constant_swab16(x) :                 \
> >>>>>            __fswab16(x))
> >>>>>    #endif
> >>>>>
> >>>>> So we can tweak selftests/bpf/bpf_endian.h and use UAPI
> >>>>> __swab16()/__swab32().
> >>>>>
> >>>>> Signed-off-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
> >>>>> ---
> >>>>>
> >>>>> v2: fixed build error, reshuffled patches (Stanislav Fomichev)
> >>>> Tested them locally with the compiler I saw the initial issues with - all
> >>>> fine, I don't see any errors with the older gcc.
> >>>>
> >>>> One last question I have is: what happens in the llvm+bpf case? Have
> >>>> you tested that? I think LLVM has all the builtins required, but since
> >>>> we are relying on the swab.h now (and it relies on
> >>>> __HAVE_BUILTIN_BSWAP16__), I wonder whether this detection works
> >>>> correctly on the llvm when targeting bpf. (sidenote: bpf_endian.h can be
> >>>> used from both userspace and bpf programs).
> >>>
> >>> Inside kernel clang compiler header (linux/compiler-clang.h) does not
> >>> define __HAVE_BUILTIN_BSWAP16__. So it will go to the "else" branch in
> >>> the above. So I think it should work with clang + bpf.
> >> Hm, isn't it the opposite of what we want then? I think for llvm+bpf we always
> >> want to use the builtins to make it properly generate
> >> BPF_TO_BE/BPF_TO_LE instructions.
> > 
> > Okay, I see. Then this patch will not achieve that.
> > The following are two common ways to compile a bpf program:
> >    - "clang -target bpf ...", maybe add macro __BPF__ somewhere
> >      to indicate builtin_bswap16 always available?
> >    - "clang <host target> ..." and then "llc -march=bpf ..."
> >      in this case, __BPF__ macro is not available and
> >      we will not be able to use builtin swap for bpf program.
> > 
> > Maybe use __clang__ macro (or gcc macro) to distinguish between clang 
> > and gcc. If it is gcc we will check builtin availability, otherwise,
> > we assume builtin always available? This not pretty though.
> 
> I think the way this should be fixed is the following: In case
> of LLVM (aka compiling BPF prog), we want the code to be as-is,
> in case if gcc is compiling the hostprog, we either want to keep
> using __builtin_bswap16() or fall-back to something else. Thus,
> I would suggest, we add a new feature test for tooling infra under
> tools/build/feature/ that compiles a dummy prog with __builtin_bswap16().
> And in the bpf_endian.h we define __bpf_ntohs(x) to __bpf_swab16(x)
> which either resolves to __builtin_bswap16() or some fallback
> implementation if not available. I don't think there should be much
> of an issue and it would follow the standard way to do it.
It's not as easy as llvm vs gcc. We can compile userland tests with
llvm/clang as well. We really need to distinguish between the target: bfp vs
non-bpf: always use builtins in bpf case and fallback to swab.h for
userland (or use feature detection, but swab.h should be enough in
theory).

Can we rely on __bpf__ define?

$ cat tmp.c
#ifdef __bpf__
#error a
#else
#error b
#endif
$ clang -c -target bpf tmp.c 
tmp.c:2:2: error: a
#error a
 ^
 1 error generated.

> 
> >>>>>    tools/testing/selftests/bpf/bpf_endian.h | 8 ++++----
> >>>>>    1 file changed, 4 insertions(+), 4 deletions(-)
> >>>>>
> >>>>> diff --git a/tools/testing/selftests/bpf/bpf_endian.h b/tools/testing/selftests/bpf/bpf_endian.h
> >>>>> index b25595ea4a78..1ed268b2002b 100644
> >>>>> --- a/tools/testing/selftests/bpf/bpf_endian.h
> >>>>> +++ b/tools/testing/selftests/bpf/bpf_endian.h
> >>>>> @@ -20,12 +20,12 @@
> >>>>>     * use different targets.
> >>>>>     */
> >>>>>    #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
> >>>>> -# define __bpf_ntohs(x)			__builtin_bswap16(x)
> >>>>> -# define __bpf_htons(x)			__builtin_bswap16(x)
> >>>>> +# define __bpf_ntohs(x)			__swab16(x)
> >>>>> +# define __bpf_htons(x)			__swab16(x)
> >>>>>    # define __bpf_constant_ntohs(x)	___constant_swab16(x)
> >>>>>    # define __bpf_constant_htons(x)	___constant_swab16(x)
> >>>>> -# define __bpf_ntohl(x)			__builtin_bswap32(x)
> >>>>> -# define __bpf_htonl(x)			__builtin_bswap32(x)
> >>>>> +# define __bpf_ntohl(x)			__swab32(x)
> >>>>> +# define __bpf_htonl(x)			__swab32(x)
> >>>>>    # define __bpf_constant_ntohl(x)	___constant_swab32(x)
> >>>>>    # define __bpf_constant_htonl(x)	___constant_swab32(x)
> >>>>>    #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
> >>>>> -- 
> >>>>> 2.21.0
> >>>>>
>