apr-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Justin Erenkrantz <jerenkra...@apache.org>
Subject Re: cvs commit: apr/test testatomic.c
Date Mon, 30 Dec 2002 00:28:17 GMT
--On Sunday, December 29, 2002 3:03 PM -0800 rbb@apache.org wrote:

> This highlights another problem with our test suite.  We can't have
> platform checks in our tests.  Having these checks completely
> invalidates the usefulness of APR.  These kinds of checks are the
> biggest reason that I haven't migrated testatomic to the new suite.
> I haven't had the time to figure out exactly what is going on yet.

Well, in this particular case, we have to go to some extremes to 
force certain schedulers (Solaris with LWPs) to actually go in 
parallel.  If that call isn't there, Solaris will execute all threads 
in serial.  When that happens, it invalidates the testatomic test, 
since we're explicitly trying to cause race conditions.

This is a side-effect of the Solaris scheduler rebalancing only on 
cancellation points (i.e. where the thread might block).  The 
testatomic code never hits a cancellation point, but normal 
applications tend to do that.  While I don't know what AIX's M:N 
thread library does, it wouldn't shock me if it has the same quirk. 
Solaris's N:N thread library (i.e. what Solaris 9 uses by default) 
shouldn't have this quirk.

I believe pthread_setconcurrency shouldn't be added to APR (in the 
past, Aaron and I have discussed this at length on this list), but 
its use should be allowed in the test suite because we have to 
provide hints because we're not a real program.  A real program would 
be extremely hard-pressed to create code that never hits a 
cancellation point.

I believe we might be able to rewrite testatomic to use condition 
variables (don't believe we had them when we wrote testatomic 
initially), but I'm still not sure if that will get around the 
concurrency problem (do a wakeup_all and hope the scheduler allocates 
enough kernel threads for all the available userspace threads).  And, 
doing a sleep() or sched_yield() doesn't help either as we have to 
get all of the user threads active at the same time.  The only 
solution that consistently works is to use pthread_setconcurrency.

Try as you might, APR can not escape the quirks of the native OS. 
This is one particular case.  -- justin

Mime
View raw message