Re: Ping[2]: [PATCH] Fix sporadic failure in tst-eintr1 test case

[Florian Weimer <fweimer at redhat dot com>]

On 04/10/2013 09:52 AM, Siddhesh Poyarekar wrote:
> On Wed, Apr 10, 2013 at 09:27:10AM +0200, Florian Weimer wrote:
> > I ran into this on current Fedora 18.
>
> I'm surprised it took so long :)
>
> > This smells like a bug in our implementation.  Can we fix this in
> > glibc?  Any pointers?
> >
> > pthread_join actually deallocating resources seems fairly important
> > to me as a quality-of-implementation issue, irrespective of what the
> > standard says.
>
> It's not a fault with pthread_join and I don't think this can be fixed
> with glibc.  The core problem here is the latency between the kernel
> notifying the pthread_join'er about the thread exit and the actual
> reaping of the thread where the latter is what reduces NPROC.  The
> test case goes from joining to spawning the new thread faster than the
> kernel is able to reap existing threads, resulting in a net increase
> in NPROC.

My concern is that if we can't fix the test case to reliably recover 
from this scenario, applications will have trouble to do so as well.  It 
is somewhat similar to deferring to garbage collection to close file 
descriptors.

I expect that freeing the kernel resources will become more and more 
expensive in relative terms, so this will eventually become visible with 
threading frameworks which do not perform thread pooling.

--
Florian Weimer / Red Hat Product Security Team