On Mon, Oct 26, 2020 at 03:45:55PM +1100, Michael Ellerman wrote:
> Vitaly Chikunov <vt@xxxxxxxxxxxx> writes:
> > Adding netdev and PowerPC maintainers JFYI.
>
> Thanks.
>
> > On Sat, Oct 24, 2020 at 11:23:19AM +0300, Dmitry V. Levin wrote:
> >> Hi,
> >>
> >> On Sat, Oct 24, 2020 at 02:06:41AM +0300, Vitaly Chikunov wrote:
> >> > Hi,
> >> >
> >> > Commit f143c11bb7b9 ("tools: bpf: Use local copy of headers including
> >> > uapi/linux/filter.h") introduces compilation issue on powerpc:
> >> >
> >> > builder@powerpc64le:~/linux$ make -C tools/bpf V=1
> >> > make: Entering directory '/usr/src/linux/tools/bpf'
> >> > gcc -Wall -O2 -D__EXPORTED_HEADERS__ -I/usr/src/linux/tools/include/uapi -I/usr/src/linux/tools/include -DDISASM_FOUR_ARGS_SIGNATURE -c -o bpf_dbg.o /usr/src/linux/tools/bpf/bpf_dbg.c
>
> Defining __EXPORTED_HEADERS__ is a hack to circumvent the checks in the
> uapi headers.
>
> So first comment is to stop doing that, although it doesn't actually fix
> this issue.
>
> >> > In file included from /usr/include/asm/sigcontext.h:14,
> >> > from /usr/include/bits/sigcontext.h:30,
> >> > from /usr/include/signal.h:291,
> >> > from /usr/src/linux/tools/bpf/bpf_dbg.c:51:
> >> > /usr/include/asm/elf.h:160:9: error: unknown type name '__vector128'
> >> > 160 | typedef __vector128 elf_vrreg_t;
> >> > | ^~~~~~~~~~~
> >> > make: *** [Makefile:67: bpf_dbg.o] Error 1
> >> > make: Leaving directory '/usr/src/linux/tools/bpf'
> >>
> >> __vector128 is defined in arch/powerpc/include/uapi/asm/types.h;
> >> while include/uapi/linux/types.h does #include <asm/types.h>,
> >> tools/include/uapi/linux/types.h doesn't, resulting to this
> >> compilation error.
> >
> > This is too puzzling to fix portably.
>
> I don't really understand how this is expected to work.
>
> We have tools/include/uapi/linux/types.h which is some sort of hand
> hacked types.h, but doesn't match the real types.h from
> include/uapi/linux.
>
> In particular the tools/include types.h doesn't include asm/types.h,
> which is why this breaks.
>
> I can build bpf_dbg if I copy the properly exported header in:
>
> $ make INSTALL_HDR_PATH=$PWD/headers headers_install
> $ cp headers/include/linux/types.h tools/include/uapi/linux/
> $ make -C tools/bpf bpf_dbg
> make: Entering directory '/home/michael/linux/tools/bpf'
>
> Auto-detecting system features:
> ... libbfd: [ on ]
> ... disassembler-four-args: [ on ]
>
> CC bpf_dbg.o
> LINK bpf_dbg
> make: Leaving directory '/home/michael/linux/tools/bpf
>
>
> I'm not sure what the proper fix is.
>
> Maybe sync the tools/include types.h with the real one?
Yeah, all f143c11bb7b9 did was sync the local copy with the result of
'headers_install', so if that's sufficient then I think that's the right
way to fix the immediate breakage.
> Or TBH I would have thought the best option is to not have
> tools/include/uapi at all, but instead just run headers_install before
> building and use the properly exported headers.
Agreed, although in some cases I suspect that kernel-internal parts are
being used.
Wiil
