Re: [PATCH v6 33/33] kselftest/riscv: kselftest for user mode cfi

[Zong Li]

On Sat, Oct 12, 2024 at 3:46 AM Deepak Gupta <debug@xxxxxxxxxxxx> wrote:
>
> On Fri, Oct 11, 2024 at 07:43:30PM +0800, Zong Li wrote:
> >On Fri, Oct 11, 2024 at 6:18 PM Mark Brown <broonie@xxxxxxxxxx> wrote:
> >>
> >> On Fri, Oct 11, 2024 at 01:44:55PM +0800, Zong Li wrote:
> >> > On Wed, Oct 9, 2024 at 7:46 AM Deepak Gupta <debug@xxxxxxxxxxxx> wrote:
> >>
> >> > > +       if (si->si_code == SEGV_CPERR) {
> >>
> >> > Hi Deepak,
> >> > I got some errors when building this test, I suppose they should be
> >> > fixed in the next version.
> >>
> >> > riscv_cfi_test.c: In function 'sigsegv_handler':
> >> > riscv_cfi_test.c:17:28: error: 'SEGV_CPERR' undeclared (first use in
> >> > this function); did you mean 'SEGV_ACCERR'?
> >> >    17 |         if (si->si_code == SEGV_CPERR) {
> >> >       |                            ^~~~~~~~~~
> >> >       |                            SEGV_ACCERR
> >> >
> >>
> >> Did you run "make headers_install" prior to building kselftest to get
> >> the current kernel's headers available for userspace builds?
> >
> >Yes, I have run "make header" and "make header_install" before
> >building the kselftest. This error happens when I cross compiled it,
> >perhaps I can help to check if it is missing some header files or
> >header search path.
>
> That's wierd.
>
> It doesn't fail for me even if I do not do `make headers_install`. But I am
> building kernel and selftests with toolchain which supports shadow stack and
> landing pad. It's defined in `siginfo.h`. When I built toolchain, I did point
> it at the latest kernel headers. May be that's the trick.
>
> """
>
> $ grep -nir SEGV_CPERR /scratch/debug/linux/kbuild/usr/include/*
> /scratch/debug/linux/kbuild/usr/include/asm-generic/siginfo.h:240:#define SEGV_CPERR    10      /* Control protection fault */
>
> $ grep -nir SEGV_CPERR /scratch/debug/open_src/sifive_cfi_toolchain/INSTALL_Sept18/sysroot/usr/*
> /scratch/debug/open_src/sifive_cfi_toolchain/INSTALL_Sept18/sysroot/usr/include/asm-generic/siginfo.h:240:#define SEGV_CPERR    10      /* Control protection fault */
> /scratch/debug/open_src/sifive_cfi_toolchain/INSTALL_Sept18/sysroot/usr/include/bits/siginfo-consts.h:139:  SEGV_CPERR                  /* Control protection fault.  */
> /scratch/debug/open_src/sifive_cfi_toolchain/INSTALL_Sept18/sysroot/usr/include/bits/siginfo-consts.h:140:#  define SEGV_CPERR  SEGV_CPERR
>
> """

In my case, because the test files don't explicitly include siginfo.h,
I assume it's expected that siginfo.h will be included through
signal.h. Regarding the header search path, it will eventually locate
signal.h in toolchain_path/sysroot/usr/include/. In my
toolchain_path/sysroot/usr/include/signal.h, it doesn't include any
signal.h; instead, signal.h will be included from
toolchain_path/sysroot/usr/include/linux/signal.h or
kernel_src/usr/include/linux/signal.h rather than
toolchain/sysroot/usr/include/signal.h. I think that is why I lost the
SEGV_CPERR definition. Is there any difference with you?

>