libpthread patch
David Xu
davidxu at freebsd.org
Sun Apr 13 20:59:03 PDT 2003
After tested Daniel's pthread patch, I found 50% of ACE test
program are core dumpped on my machine. So I studied the
libpthread source code after applied the patch, I found that
the main problem is thread state transition is not atomic,
for example in thr_mutex:
mutex_queue_enq(*m, curthread);
curthread->data.mutex = *m;
/*
This thread is active and is in a critical
region (holding the mutex lock); we should
be able to safely set the state.
*/
THR_SET_STATE(curthread, PS_MUTEX_WAIT);
/* Unlock the mutex structure: */
THR_LOCK_RELEASE(curthread, &(*m)->m_lock);
/* Schedule the next thread: */
_thr_sched_switch(curthread);
thread sets its state to PS_MUTEX_WAIT, and call _thr_sched_switch,
but it is not under scheduler lock, so there is a race between
THR_SET_STATE and thr_sched_switch.
I have inserted _kse_critical_enter() before THR_SET_STATE,
the code looks as following:
mutex_queue_enq(*m, curthread);
curthread->data.mutex = *m;
_kse_critical_enter();
/*
This thread is active and is in a critical
region (holding the mutex lock); we should
be able to safely set the state.
*/
THR_SET_STATE(curthread, PS_MUTEX_WAIT);
/* Unlock the mutex structure: */
THR_LOCK_RELEASE(curthread, &(*m)->m_lock);
/* Schedule the next thread: */
_thr_sched_switch(curthread);
I also commented out most code in thr_lock_wait() and
thr_lock_wakeup(), I think without better scheduler lock,
these code has race condition, and in most case will
this cause a thread be reinserted into runq while it
is already in this queue.
now, I can run ACE test programs without any core dumpped,
and only the following program are failed:
Cached_Conn_Test
Conn_Test
MT_Reactor_Timer_Test
Malloc_Test
Process_Strategy_Test
Thread_Pool_Test
a complete log file is at:
http://people.freebsd.org/~davidxu/run_test.log
the libpthread package I modified is at:
http://people.freebsd.org/~davidxu/libpthread.tgz
Also, I can run crew test program without any problem.
I think the whole scheduler lock should be reworked
to allow state transition is in atomic, my change is
not SMP safe, only works on UP, because kse_critical_enter
is only works for UP system. If we fixed this scheduler lock
problem, I think the libpthread will be stable enough.
David Xu
More information about the freebsd-threads
mailing list