[Bug 20232] New: mciwave breaks on MSDN example

http://bugs.winehq.org/show_bug.cgi?id=20232

Summary: mciwave breaks on MSDN example
           Product: Wine
           Version: 1.1.30
          Platform: All
        OS/Version: All
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: winmm&mci
        AssignedTo: [e-mail]
        ReportedBy: [e-mail]

I'm reviewing mciwave and mciseq and found numerous issues.  Concurrently, I'm
augmenting the MCI testsuite.  I've attached my current mciwave
work-in-progress tests to give you an early insight into my work.

The tests so far confirm my model of MCI: I view it like an old-fashioned tape
deck with buttons, which you can press in any order.  Still there's error
checking, e.g. you can't resume when stopped.  Like a tape deck, you can switch
from playing to recording at any time.
Oddly, you can combine fast forward and play in one command, but not rewind
with play.

The attached tests work fine on (at least one machine with) MS-Windows, but
cause Wine to fail and hang in multiple ways. You've been warned!

I'd be pleased if people could run the tests on more instances of MS-Windows.

I plan to strip down this test file until it at least does not hang anymore in
Wine.  Then only can it be submitted.  Concurrently, I can submit patches to
the mci* codebase.

Current issues with the mci code in Wine are:
 - not distinguishing between STOPPING and STOPPED, PLAYING and GOING TO PLAY,
etc.;
 - seriously broken asynchronous execution;
 - non-error-proof use of InterlockDecrement. It must only be called
   when waveInAddBuffer and waveOutWrite succeed, as those cause
   callbacks to happen;
 - RIFF .wav file not always correctly written;
 - the mmio is as much as resource as the wave device and must be
   properly released (mmioAscend, not only when saving);
 - many items here and there:
  + bogus 44000Hz frequency;
  + premature return (in mcicda);
  + conversion between #bytes and samples;
  + switching fInput while playing;
  + copy&paste errors (InterlockedDecrement need be initialised
    differently when recording and playing)

The major issue is concurrency. There I'm not sure how to proceed and rewrite
mciwave (and mciseq and mcicda).

o One model, Erlang-like, which the OSS driver also implements, is one thread
per play and exclusively using message passing to receive commands.  This would
simplify NOTIFY everywhere.  It clearly offers the advantage of being closest
to the sequential execution model that is easy to reason about.  IMHO, it does
not play nicely with the pause command (state machine).

o Continue as currently written, but then think twice and even 3-7 times about
how to perform locking among the concurrent threads:
 - Put InterlockIncrement() to more uses?
 - Put Events to more uses?
 - Unlike mciavi, there's no CriticalSection in mciwave and mcicda, yet
   I'm unconvinced that the current synchronisation via volatile dwStatus
   can suffice.
 - But then, identifying the critical sections will be tedious and error prone.

o Combinations of both, e.g. one thread per play or record (not unlike the
present code), and dealing (correctly) with PAUSE, STOP and RESUME inside it
(via e.g. events to restart a paused thread, or indirectly, by relying on the
callbacks+events sent when invoking waveOutRestart and waveInStart -- is it
legal at all to call those from a different thread)?

o Is there any other concurrency model suitable in Wine?

Regards,
 Jörg Höhle