Re: [patch] Fix for gdb.threads/staticthreads.exp failure on Linux

[Doug Evans <dje@google.com>]

On Mon, Jun 22, 2009 at 1:21 PM, Paul Pluzhnikov<ppluzhnikov@google.com> wrote:
> On Mon, Jun 22, 2009 at 11:09 AM, Michael Snyder<msnyder@vmware.com> wrote:
>
>> Have you, or could you, run regression testing in i686?
>
> This is glibc-version related, rather than x86_64 vs. i686:
>
> Using glibc-2.7: "runtest gdb.threads/staticthreads.exp" gives:
>
>  # of expected passes            8
>  # of known failures             1
>
> in both i686 and x86_64 modes.
>
> But using glibc-2.3.6 (which is my target):
>
>  FAIL: gdb.threads/staticthreads.exp: running to main in runto
>  FAIL: gdb.threads/staticthreads.exp: Continue to main's call of sem_post
>  FAIL: gdb.threads/staticthreads.exp: handle SIG32 helps
>
>                  === gdb Summary ===
>
>  # of expected passes            6
>  # of unexpected failures        3
>  # of known failures             1
>
> again in both i686 and x86_64 modes.
>
> Thanks,
> --
> Paul Pluzhnikov
>

I looked into this a bit.

I think(!) this is the glibc patch that fixes things.

2007-05-16  Roland McGrath  <roland@redhat.com>

        * td_ta_thr_iter.c (iterate_thread_list): Make FAKE_EMPTY bool.
        Use th_unique=0 in fake descriptor before initialization.

Regardless, this is indeed glibc specific, and older glibcs fail while
newer ones pass.
I like the patch.  If thread enumeration fails during
try_thread_db_load_1 we ignore it and leave gdb to try again later.
When we get to this point we've already verified libthread_db is
(mostly) happy.
I kinda wonder, though, if that works then maybe gdb shouldn't do
thread enumeration at all here.
Running the testcase with --target_board=native-gdbserver doesn't have
this problem.  Some research as to what gdbserver is doing would be
illuminating.  "consistency is good": can you research what's
happening in gdbserver that it works there, and see if it's reasonable
to do the same thing in gdb?

[digression: GDB does use glibc's support for determining the glibc
version, but here we want the glibc version of libthread_db so that's
out.  Hmmm, though it looks like the existing use of
gnu_get_libc_version is to get the libthread_db version which can be
different now that we have libthread-db-search-path.  ...  This seems
like a problem, though in practice I suspect it's ok.  [grep for
gnu_get_libc_version in linux-thread-db.c]]

I think a comment needs to be added to your patch, probably to the
call site of thread_db_find_new_threads_silently, that thoroughly
explains what's going on (assuming that's still the best solution).

btw, the reason for the 1 known failure is, I'm guessing, because gdb
was built with the same toolchain that uses the older glibc, but the
test was run using a native toolchain that uses a newer glibc and the
mismatch is sufficient to trigger the failure.  At least that explains
things in my sandbox.  If I rebuild gdb with native gcc, or set
libthread-db-search-path to find the newer glibc, the kfail goes away.