Refactored tests in clang to use barrier_init/wait:
http://llvm.org/viewvc/llvm-project?view=revision&revision=226659
There are still few sleep call, like when we need to wait for a thread
to exit (there is really no point to insert barrier_wait); or when we
need to wait for a thread to _block_ inside of pthread_cond_wait; or
when a thread needs to process signals. But all remaining uses of
sleep are commented.
On Wed, Jan 21, 2015 at 11:58 AM, Dmitry Vyukov <dvyu...@google.com> wrote:
> On Tue, Jan 20, 2015 at 10:09 AM, Bernd Edlinger
> <bernd.edlin...@hotmail.de> wrote:
>>
>> Hi,
>>
>> On Mon, 19 Jan 2015 18:49:21, Konstantin Serebryany wrote:
>>>
>>> [text-only]
>>>
>>> On Mon, Jan 19, 2015 at 7:42 AM, Mike Stump <mikest...@comcast.net> wrote:
>>>> On Jan 19, 2015, at 12:43 AM, Dmitry Vyukov <dvyu...@google.com> wrote:
>>>>> I can't really make my mind on this. I would mildly prefer sleep's (if
>>>>> they work reliably!).
>>>>
>>>> Let me state it more forcefully.
>>> You don't have to convince us here.
>>> I'd love to get rid of sleep calls in the tsan test suite -- they are
>>> a minor but a constant annoyance.
>>> But I also want to keep the tests *very simple*, i.e.
>>> 1. Single file w/o any non-system includes, no linking of extra
>>> libraries/objects
>>> 2. Not too much extra code. (ideally, 1 line for init, 1 line for
>>> "signal", 1 line for "wait")
>>> 3. Strictly posix or c++11 (unless we are testing something specific)
>>>
>>> Your idea with barrier_wait/dlsym sounds interesting, but I can't see
>>> the code in this mail thread.
>>> What do I miss?
>>>
>>
>> We discussed two alternatives to sleep:
>>
>> 1. step function, optionally with sched_yield to make it somewhat less busy
>> waiting:
>> __attribute__((no_sanitize_thread))
>> void step (int i)
>> {
>> while (__atomic_load_n (&serial, __ATOMIC_ACQUIRE) != i - 1)
>> sched_yield();
>> __atomic_store_n (&serial, i, __ATOMIC_RELEASE);
>> }
>> 2. tsan-invisible barriers:
>>
>> cat tsan_barrier.h
>> /* TSAN-invisible barriers. Link with -ldl. */
>> #include <pthread.h>
>> #include <dlfcn.h>
>>
>> static __typeof(pthread_barrier_wait) *barrier_wait;
>>
>> static
>> void barrier_init (pthread_barrier_t *barrier, unsigned count)
>> {
>> void *h = dlopen ("libpthread.so.0", RTLD_LAZY);
>> barrier_wait = (__typeof (pthread_barrier_wait) *)
>> dlsym (h, "pthread_barrier_wait");
>> pthread_barrier_init (barrier, NULL, count);
>> }
>>
>>
>> We preferred the second alternative, because it does not do busy waiting.
>> We include this header file in every positive test case and link with -ldl.
>
>
> The step approach looks better to me at first sight.
>
> Busy waiting looks like a weak argument in this context. It's
> absolutely non performance-critical and a yield or usleep(10) will
> solve it more than sufficiently.
>
> I will check how complex to make its implementation invisible to tsan
> (I suspect that clang does not ignore atomic operations when
> no_sanitize_thread attribute is given) and whether it actually makes
> more complex tests simpler to write (as compared to the barrier
> approach).
