Re: selftests: sigaltstack: sas # exit=1 - # Bail out! SP is not on sigaltstack - on clang build

[Nick Desaulniers]

On Wed, Mar 8, 2023 at 8:23 AM Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>
> On Tue, Mar 7, 2023 at 7:14 PM Naresh Kamboju <naresh.kamboju@xxxxxxxxxx> wrote:
> >
> > + LLVM
>
> The offending code seems to be:
>
> #if __s390x__
>         register unsigned long sp asm("%15");
> #else
>         register unsigned long sp asm("sp");
> #endif
>
>         if (sp < (unsigned long)sstack ||
>                         sp >= (unsigned long)sstack + stack_size) {
>                 ksft_exit_fail_msg("SP is not on sigaltstack\n");
>         }
>
> Is that actually expected to work?  asm("sp") is a horrible hack.  I
> would, maybe naively, expect a compiler to analyze this code, think
> "sp is unconditionally uninitialized", and treat the comparison as
> always-UB and thus generate whatever code seems convenient.

Spot-on. -Wuninitialized should warn about that.
https://godbolt.org/z/do9Kqa3cG

Kees mentioned we should be using `current_stack_pointer`. I'll whip
up a patch using that.

>
> --Andy
>
> >
> > On Wed, 8 Mar 2023 at 00:58, Chang S. Bae <chang.seok.bae@xxxxxxxxx> wrote:
> > >
> > > On 3/6/2023 10:57 PM, Naresh Kamboju wrote:
> > > > kselftest: sigaltstack built with clang-16 getting failed but passed with
> > > > gcc-12 build. Please find more details about test logs on clang-16 and
> > > > gcc-12 and steps to reproduce locally on your machine by using tuxrun.
> > > >
> > > > Reported-by: Linux Kernel Functional Testing <lkft@xxxxxxxxxx>
> > > >
> > > > Test log:
> > > > ----------
> > > >
> > > > Linux version 6.3.0-rc1-next-20230307 (tuxmake@tuxmake) (Debian clang
> > > > version 16.0.0 (++20230228093516+60692a66ced6-1~exp1~20230228093525.41),
> > > > Debian LLD 16.0.0) #1 SMP PREEMPT @1678159722
> > > > ...
> > > > kselftest: Running tests in sigaltstack
> > > > TAP version 13
> > > > 1..1
> > > > # selftests: sigaltstack: sas
> > > > # # [NOTE] the stack size is 21104
> > > > # TAP version 13
> > > > # 1..3
> > > > # ok 1 Initial sigaltstack state was SS_DISABLE
> > > > # Bail out! SP is not on sigaltstack
> > > > # # Planned tests != run tests (3 != 1)
> > > > # # Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
> > > > not ok 1 selftests: sigaltstack: sas # exit=1
> > > <snip>
> > >
> > > > Linux version 6.3.0-rc1-next-20230307 (tuxmake@tuxmake)
> > > > (aarch64-linux-gnu-gcc (Debian 12.2.0-14) 12.2.0, GNU ld (GNU Binutils
> > > > for Debian) 2.40) #1 SMP PREEMPT @1678159736
> > > > ...
> > > > kselftest: Running tests in sigaltstack
> > > > TAP version 13
> > > > 1..1
> > > > # selftests: sigaltstack: sas
> > > > # # [NOTE] the stack size is 50080
> > > > # TAP version 13
> > > > # 1..3
> > > > # ok 1 Initial sigaltstack state was SS_DISABLE
> > > > # # [RUN] signal USR1
> > > > # ok 2 sigaltstack is disabled in sighandler
> > > > # # [RUN] switched to user ctx
> > > > # # [RUN] signal USR2
> > > > # # [OK] Stack preserved
> > > > # ok 3 sigaltstack is still SS_AUTODISARM after signal
> > > > # # Totals: pass:3 fail:0 xfail:0 xpass:0 skip:0 error:0
> > > > ok 1 selftests: sigaltstack: sas
> > >
> > > At glance, the log shows the altstack size difference between LLVM and GCC.
> > >
> > > But, when I tried with the LLVM that I have,
> > >
> > >      $ clang --version
> > >      clang version 13.0.0 ...
> > >
> > > it failed only with this compiler:
> > >
> > >      $ rm sas;clang -o sas sas.c;./sas
> > >      # [NOTE]        the stack size is 8192
> > >      TAP version 13
> > >      1..3
> > >      ok 1 Initial sigaltstack state was SS_DISABLE
> > >      Bail out! SP is not on sigaltstack
> > >      # Planned tests != run tests (3 != 1)
> > >      # Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
> > >
> > >      $ rm sas;gcc -o sas sas.c;./sas
> > >      # [NOTE]        the stack size is 8192
> > >      TAP version 13
> > >      1..3
> > >      ok 1 Initial sigaltstack state was SS_DISABLE
> > >      # [RUN] signal USR1
> > >      ok 2 sigaltstack is disabled in sighandler
> > >      # [RUN] switched to user ctx
> > >      # [RUN] signal USR2
> > >      # [OK]  Stack preserved
> > >      ok 3 sigaltstack is still SS_AUTODISARM after signal
> > >      # Totals: pass:3 fail:0 xfail:0 xpass:0 skip:0 error:0
> > >
> > > The same is true with some old versions -- e.g. the one that came with
> > > commit 0c49ad415512 ("tools/testing/selftests/sigaltstack/sas.c: improve
> > > output of sigaltstack testcase"):
> > >
> > >      $ rm sas;clang -o sas sas.c;./sas
> > >      [OK]    Initial sigaltstack state was SS_DISABLE
> > >      [FAIL]  SP is not on sigaltstack
> > >
> > >      $ rm sas;gcc -o sas sas.c;./sas
> > >      [OK]    Initial sigaltstack state was SS_DISABLE
> > >      [RUN]   signal USR1
> > >      [OK]    sigaltstack is disabled in sighandler
> > >      [RUN]   switched to user ctx
> > >      [RUN]   signal USR2
> > >      [OK]    Stack preserved
> > >      [OK]    sigaltstack is still SS_AUTODISARM after signal
> > >      [OK]    Test passed
> > >
> > > So, this test failure appears to have been there for a while. I think
> > > the LLVM folks need to take a look at it.
> > >
> > > Thanks,
> > > Chang
>


-- 
Thanks,
~Nick Desaulniers