Re: ATF tests still failing to complete within 2 hours on amd64

[Thor Lancelot Simon <tls%panix.com@localhost>]

On Fri, Nov 25, 2011 at 09:19:04AM +0200, Andreas Gustafsson wrote:
> Paul Goyette wrote:
> > A couple of network-related tests now seem to require more than five 
> > minutes per test case, and I'm suspecting that there are others that are 
> > taking significantly more time than they used to.  A complete test run 
> > used to take about 100 minutes - my last successful complete pass was a 
> > few days ago, on 2011-11-19 at 18:50:05 UTC
> 
> I believe these are the same networking-related tests that the TNF
> i386 test server shows breaking at the time of Thor's overhaul of the
> random number subsystem:
> 
>   
> http://releng.netbsd.org/b5reports/i386/commits-2011.11.html#2011.11.19.20.44.58
> 
> I notified Thor about these failures by email on the 20th.

Indeed he did.  I've been away, with intermittent network connectivity,
but I have been trying to understand what's going on with those tests,
and failing, so far.

When I run them as individual test cases rather than in sequence, some
of them work fine; then others fail because rump processes from earlier
cases are dangling.  Others do seem to take a very long time but finally
complete.

I don't see even remotely similar behavior doing the same operations with
a real kernel, so I'm sure I broke rump somehow.  But the combination
of rump's awful build system (which already had the arc4random and rnd code
working considerably differently before I touched it, a problem I made
worse because I had to add more stubs) and the problems caused for rump
by the wrong separation of code in rnd itself -- rump *really* does not like
having the pseudodevice and core kernel functionality in the same source file
-- mean it's not easy to understand what's going on, nor to experiment to try
to fix it.

I'll figure it out, but I can't promise it will be fast.

Thor