From: John Bartholomew
Subject: [SDL] SDL Semaphore implementation broken on Windows?
Hi,
Over the past couple of days, I've been battling with SDL, SDL_Mixer and SMPEG to try to find an audio hang bug. I believe I've found the problem, which I think is a race condition inside SDL's semaphore implementation (at least the Windows implementation). The semaphore code uses Windows' built in semaphore functions, but it also maintains a separate count value. This count value is updated with bare increment and decrement operations in SemPost and SemWaitTimeout - no locking primitives to protect them.
In tracking down the apparent audio bug, I found that at some point a semaphore's count value was being decremented to -1, which is clearly not a valid value for it to take.
I'm still not certain exactly what sequence of operations is occuring for this to happen, but I believe that overall it's a race condition between a thread calling SemPost (which increments the count) and the thread on the other end calling SemWait (which decrements it).
I will try to make a test case to verify this, but I'm not sure if I'll be able to (threading errors being difficult to reproduce even in the best circumstances).
However, assuming this is the cause of my problems, there is a very
simple fix:
Windows provides InterlockedIncrement() and InterlockedDecrement()
functions to perform increments and decrements which are guaranteed to be atomic. So the fix is in thread/win32/SDL_syssem.c: replace occurrences of --sem->count with InterlockedDecrement(&sem->count); and replace occurrences of ++sem->count with InterlockedIncrement(&sem->count);
This is using SDL v1.2.12, built with VC++ 2008 Express, running on a
Core 2 duo processor.
Description From Michael Stone 2008-09-25 19:27:29 (-) [reply]
To determine whether a pid is occupied with the kill(pid, 0) idiom, you have to
test
#include <signal.h>
#include <errno.h>
kill(pid, 0) < 0 && errno == ESRCH
not just
#include <signal.h>
kill(pid, 0) < 0
otherwise you get incorrect results when pid is running as a different user
(causing kill(pid, 0) to return -1 + EPERM).
src/audio/alsa/SDL_alsa_audio.c is certainly affected by this bug in both
1.2.13 and 1.3-trunk. It probably occurs in other places as well.
#ifndef SDL_VIDEO_DRIVER_X11_DYNAMIC_XSS
#define SDL_VIDEO_DRIVER_X11_DYNAMIC_XSS NULL
#endif
so that the X code would compile. It looked like it was needed
Instead of the GNOME screensaver respecting the MIT-SCREEN-SAVER extension
and registering for the screensaver idle event, it uses a completely different
control mechanism. Our alternatives are to use dbus (very complicated) or
this really painful hack.
- use SDL_getenv, not getenv ...
- no more support for 0.9.25 - not even mentioned any longer on directfb.org
- fix fullscreen issues
- add a simple window manager unless the directfb team comes up with a working wm.
The driver has support for a very, very basic window manager you may
want to use when runnning with "wm=default". Use
export SDL_DIRECTFB_WM=1
to enable basic window borders including icon support. In order to have the window title rendered,
you need to have the following font installed:
/usr/share/fonts/truetype/freefont/FreeSans.ttf
* Some math functions become intrinsic in release mode, so we need to
convert all the math functions into SDL math functions, like we did
with the stdlib functions.
* Constant initializers of 8-bit values become calls to memset() in
release mode, but memset() itself is an intrinsic when explicitly
called. So we'll just explicitly call memset() in those cases.
