Re: [PATCH bpf-next] selftests/bpf: Fix building error when using userspace pt_regs

[Andrii Nakryiko]

On Mon, Dec 20, 2021 at 4:58 PM Pu Lehui <pulehui@xxxxxxxxxx> wrote:
>
>
>
> On 2021/12/20 22:02, Pu Lehui wrote:
> >
> >
> > On 2021/12/18 0:45, Andrii Nakryiko wrote:
> >> On Thu, Dec 16, 2021 at 6:25 PM Pu Lehui <pulehui@xxxxxxxxxx> wrote:
> >>>
> >>>
> >>>
> >>> On 2021/12/16 12:06, Andrii Nakryiko wrote:
> >>>> On Tue, Dec 14, 2021 at 5:54 AM Pu Lehui <pulehui@xxxxxxxxxx> wrote:
> >>>>>
> >>>>> When building bpf selftests on arm64, the following error will occur:
> >>>>>
> >>>>> progs/loop2.c:20:7: error: incomplete definition of type 'struct
> >>>>> user_pt_regs'
> >>>>>
> >>>>> Some archs, like arm64 and riscv, use userspace pt_regs in
> >>>>> bpf_tracing.h, which causes build failure when bpf prog use
> >>>>> macro in bpf_tracing.h. So let's use vmlinux.h directly.
> >>>>
> >>>> We could probably also extend bpf_tracing.h to work with
> >>>> kernel-defined pt_regs, just like we do for x86 (see __KERNEL__ and
> >>>> __VMLINUX_H__ checks). It's more work, but will benefit other end
> >>>> users, not just selftests.
> >>>>
> >>> It might change a lot. We can use header file directory generated by
> >>> "make headers_install" to fix it.
> >>
> >> We don't have dependency on "make headers_install" and I'd rather not
> >> add it.
> >>
> >> What do you mean by "change a lot"?
> >>
> > Maybe I misunderstood your advice. Your suggestion might be to extend
> > bpf_tracing.h to kernel-space pt_regs, while some archs, like arm64,

yes

> > only support user-space. So the patch might be like this:
> >
> > diff --git a/tools/lib/bpf/bpf_tracing.h b/tools/lib/bpf/bpf_tracing.h
> > index db05a5937105..2c3cb8e9ae92 100644
> > --- a/tools/lib/bpf/bpf_tracing.h
> > +++ b/tools/lib/bpf/bpf_tracing.h
> > @@ -195,9 +195,13 @@ struct pt_regs;
> >
> >   #elif defined(bpf_target_arm64)
> >
> > -struct pt_regs;
> > +#if defined(__KERNEL__)
> > +#define PT_REGS_ARM64 const volatile struct pt_regs
> > +#else
> >   /* arm64 provides struct user_pt_regs instead of struct pt_regs to
> > userspace */
> >   #define PT_REGS_ARM64 const volatile struct user_pt_regs
> > +#endif
> > +
> >   #define PT_REGS_PARM1(x) (((PT_REGS_ARM64 *)(x))->regs[0])
> >   #define PT_REGS_PARM2(x) (((PT_REGS_ARM64 *)(x))->regs[1])
> >   #define PT_REGS_PARM3(x) (((PT_REGS_ARM64 *)(x))->regs[2])
> >
> Please ignore the last reply. User-space pt_regs of arm64/s390 is the
> first part of the kernel-space's, it should has covered both kernel and
> userspace.

Alright, so is there still a problem or not? Looking at the definition
of struct pt_regs for arm64, just casting struct pt_regs to struct
user_pt_regs will indeed just work. So in that case, what was your
original issue?

> >>>
> >>> --- a/tools/testing/selftests/bpf/Makefile
> >>> +++ b/tools/testing/selftests/bpf/Makefile
> >>> @@ -294,7 +294,8 @@ MENDIAN=$(if
> >>> $(IS_LITTLE_ENDIAN),-mlittle-endian,-mbig-endian)
> >>>    CLANG_SYS_INCLUDES = $(call get_sys_includes,$(CLANG))
> >>>    BPF_CFLAGS = -g -D__TARGET_ARCH_$(SRCARCH) $(MENDIAN) \
> >>>               -I$(INCLUDE_DIR) -I$(CURDIR) -I$(APIDIR) \
> >>> -            -I$(abspath $(OUTPUT)/../usr/include)
> >>> +            -I$(abspath $(OUTPUT)/../usr/include) \
> >>> +            -I../../../../usr/include
> >>>>>
> >>>>> Signed-off-by: Pu Lehui <pulehui@xxxxxxxxxx>
> >>>>> ---
> >>>>>    tools/testing/selftests/bpf/progs/loop1.c     |  8 ++------
> >>>>>    tools/testing/selftests/bpf/progs/loop2.c     |  8 ++------
> >>>>>    tools/testing/selftests/bpf/progs/loop3.c     |  8 ++------
> >>>>>    tools/testing/selftests/bpf/progs/loop6.c     | 20
> >>>>> ++++++-------------
> >>>>>    .../selftests/bpf/progs/test_overhead.c       |  8 ++------
> >>>>>    .../selftests/bpf/progs/test_probe_user.c     |  6 +-----
> >>>>>    6 files changed, 15 insertions(+), 43 deletions(-)
> >>>>>
> >>>>
> >>>> [...]
> >>>> .
> >>>>
> >> .
> >>
> > .