Recent patches introduced a regression error

[Paul Millar]

On Thu, 1 Aug 2002, [iso-8859-1] Sylvain Petreolle wrote:
> http://www.winehq.com/hypermail/wine-devel/2002/07/0386.html

Hi Sylvain,

Sorry, I meant to reply to that email earlier. I've notice the same or
similar thing with WRT. If you look at around Friday July 19 (see Win3.0
tests column in [1]), you'll see four tests weren't conducted. Next build
cycle they were.

The same thing happened earlier, although less clearly, around 5th July
(see [2]).  On Friday, new regression tests were added, two of which
failed straight away (the C:, D: problem). Because the tests are only
known by their file-name and line-number, the new tests cause the "-27
tests +42 tests" on each build. Except, for Win2.0 tests 31 tests were not
conducted, ie 4 extra test. The following Monday, these tests were
conducted, but 4 tests were not conducted for the Win3.1 platform, only
for those tests to be conducted on Tuesday (+59 tests for Win3.1 instead
of +55).

What happened was this: the patches tend to be applied during the night
(local time) and unless I happen to be working late, I discover the
effects of the patches in the morning.  For those particular builds I
discovered that the regression tests had frozen, just as you describe. I
had to kill wine so the cycle could complete. With a bit of detective
work, I traced the problem down to wininet.  It seemed that the exact
order in which the tests are run is random (multiple threads? -- I haven't
had a chance to check the code yet).

As the problem didn't immediately resurface after Monday 8th July, I
assumed that it was a low-probability thread contention/deadlock issue. I
tried conducting just the regression tests for wininet, without being able
to reproducing the problem, but I didn't think to try the whole tests
(just assumed that the tests were independent).

It looks like something has changed, so the occasional-hard-lock problem
has been replaced by the system freezing for many minutes before reporting
a regression test failure.

So, Sylvain, when you run the tests now, does it eventually time-out
returning a error, or does it freeze forever?

Cheers,

Paul

[1] http://www.astro.gla.ac.uk/users/paulm/WRT/wrt.php?max=5&start=102
[2] http://www.astro.gla.ac.uk/users/paulm/WRT/wrt.php?max=5&start=94

----
Paul Millar


>  --- Paul Millar <paulm at astro.gla.ac.uk> a écrit : > Hi,
> > 
> > According to WRT [1], one of todays patches caused a regression
> > error,
> > picked up by one of the wininet DLL tests. Only, none of the patches
> > seem
> > to have touched wininet.
> > 
> > The problem is with the tests at dlls/wininet/tests/http.c lines 117
> > and 
> > 157. The test freezes for up to 4 minutes, then returns the failures.
> > 
> > As an act in desperation, I've tried reversing some of the patches
> > (1-4,
> > 6, and 7 from the WRT list) without success.
> > 
> > I'm also a little suspicious that this might be something to do with
> > Glasgow (or me :)  I've been able to reproduce the problem on two
> > other
> > (local) machines, but I've not tried on any remote machines. I've
> > also
> > tried using earlier cvs versions (from a few days ago), but that
> > build
> > fails the wininet tests too. It would be good if someone could
> > confirm the
> > problem.
> > 
> > Sorry if this turns out to be a red herring ...
> > 
> > Paul.
> > 
> > PS It's quite late here, so I might not be able to reply until
> > tomorrow.
> > 
> > [1] http://www.astro.gla.ac.uk/users/paulm/WRT/wrt.php