Re: Fix race in tst-mqueue5

["Paul E. Murphy" <murphyp at linux dot vnet dot ibm dot com>]

On 01/12/2016 07:58 PM, Carlos O'Donell wrote:
> On 01/12/2016 02:23 PM, Paul E. Murphy wrote:
>> This seems to fix the test for ppc, and probably others too.
> 
> Had you considered the 'sleep (1);' on line 394 as another source of problems?

That is yet another problem, though seemingly less likely. I ran the test in a loop
on a relatively busy machine for a night and didn't catch any failures. I'm not sure
how or if it can be fixed.

> Must it be the case that by the time '(void) pthread_barrier_wait (b3);' returns,
> the SIGRTMIN has been delivered and handled? Why isn't 'sigwait' required to make
> this test operate without a race bewtween checking 'rtmin_cnt != 2' and the signal
> arriving and being handled?

I think there is an assumption that calling mqrecv() in do_test() will generate a
signal in the child process when the child is in one of the following states:

1. thr() blocked on barrier b3, do_child() is still approaching barrier b3.
2. do_child() blocked on barrier b3, thr() is still approaching barrier b3.
3. Both are waiting in barrier b3.

Masking the signal on do_child() guarantees thr() will handle the exception.
Furthermore, this assumes the signal to thr() cannot be delayed or blocked. I don't
know if that is a safe assumption. Though, the assumption is made elsewhere in the
test.

> While I agree that any fix that makes tst-mqueue5 fail less spuriously is a good
> thing, I'm curious about your review of the test as a whole (now that I've looked
> at it again).

Each call to pthread_barrier_wait should have a comment to quickly match it up
with the matching call in the other one or two functions :).

It does validate the documented behavior of mq, so it's a good test. If the test
is still troublesome, maybe the signals can be replaced with another means of
validating the asynchronous API.

Thanks,
Paul