[LTP] [PATCH v2] pthread_detach: avoid threads sharing stack

[Jan S. <jst...@re...>]

This testcase was observed to crash with segmentation fault:
http://www.mail-archive.com/ltp...@li.../msg14692.html

Following issues have been found:
- When detached thread signals it's done, it's not really done yet
  and it is possible for scheduler to switch to test() thread and create
  another thread(s) that uses same altstack. sem_post/sem_wait
  is not reliable way to tell when detached thread is finished.
  test() thread can continue to spawn new threads and unprivileged user
  can run out of resources.
- Threads which fire SIGUSR signals can do so very frequently,
  so threaded() thread may not progress at all and it starts to
  progress only when signal threads are told to stop sending signals.

This patch modifies testcase in following way:

Before
-------
test() loops over all scenarios repeatedly and spawns new threaded()
as soon as previous one called sem_post.

main +
     |\
     | send_sig1 (thread)
     |\
     | send_sig2 (thread)
     \
       test (thread)
         \
          threaded (detached thread)

After
------
Only one thread is sending signals (SIGUSR1) with gradually increasing
sleep time to make sure threaded() can progress. test() loops over
scenarios just once.

main() loops for specified time, it spawns new child and waits for it
to finish. This will avoid using same stack by 2+ threads and there will be
no more than NSCENAR threads at a time.

main (procA)
             \
              do_child (procB) +
                               |\
                               | send_sig (thread)
                                \
                                 test (thread)
                                   \
                                    threaded (detached thread)

Modified version have been succesfully tested by reporter:
routerstation(mips), x86_64(ubuntu11.10 & redhat 5.3)
and x86(redhat 5.3) each 100+ times

and by me: 100+ times as root and unprivileged user on:
RHEL5.8 (i386, x86_64, ia64, ppc64, s390x)
RHEL6.0 (i386, x86_64, ppc64, s390x)
RHEL6.3 (i386, x86_64, ppc64, s390x)

Following glibc bug has been observed during testing on ppc64:
Bug 852445 - gcc_assert failure in _Unwind_SetSpColumn() on ppc64
https://bugzilla.redhat.com/show_bug.cgi?id=852445

This patch is currently not making any workarounds to avoid
hitting this issue, because so far it is known to happen only
rarely on ppc64 when unprivileged user runs the testcase.
If this issue becomes more significant, testcase can be modified
to skip scenarios with explicit scheduling if they are run by
unprivileged user.

Reported-by: Kang Kai <kai...@wi...>
Signed-off-by: Jan Stancek <jst...@re...>
---
 .../conformance/interfaces/pthread_detach/4-3.c    |  207 ++++++++++++--------
 1 files changed, 126 insertions(+), 81 deletions(-)