Mattias Helsing
2012-12-15 15:00:28 UTC
Hi all
I wanted to bring this up here before posting on osg-submissions.
I have been bit for the third time now with using OpenThreads::Mutex as a
simple synchronization point between threads. My usage is typically that I
need some functionality to run in parallell to the frame loop but I need to
tell it when to start the next loop from the frame thread. So I derive from
OpenThreads::Thread and have a OpenThreads::Mutex as a member of my new
thread class, then have a lock at the start of the parallell thread's main
loop waiting for a mutex to be released.
This works great for a week so until I test my app on windows and things
start to misbehave... Because on linux, with pthreads as the OpenThreads
backend, a lock is a lock is a lock and it doesn't matter from which thread
somebody tries to take a new lock, i.e. mutex is not recursive. While on
windows with CriticalSection (which is the Mutex backend on win32) the
thread that has the lock can relock it as many times as it likes, and from
what I read this by design and an important part of how Windows works.
Already ranting...will try to come to point.
All the above is ok, except that OpenThreads::Mutex works differently on
win32 and linux. Now - I have an implementation using Semaphores on windows
that makes Mutex work the same on win32 as with phtreads. My moves the
current behaviour to the OpenThreads::ReentrantMutex (which is currently
not implemented on win32). However - there is also some code in OpenThreads
that can only be enabled by manually defining USE_CRITICAL_SECTION that
leads me to believe that someone is already working on this or a similar
problem.
I'm getting more and more experienced in finding and fixing thread
synchronization problems (and want to avoid it as much as I can ;) , but
can't say I know everything about threads, so If you feel the urge to tell
me that i'm using the Mutex class in a bad way then please teach me bwana.
BUT the bottom line is still that the OpenThreads::Mutex class works
differerently on win32 and pthreads. We should fix it IMHO.
Avaiting your input or will post to osg-submissions (yes, it's a threat :)
/Mattias
I wanted to bring this up here before posting on osg-submissions.
I have been bit for the third time now with using OpenThreads::Mutex as a
simple synchronization point between threads. My usage is typically that I
need some functionality to run in parallell to the frame loop but I need to
tell it when to start the next loop from the frame thread. So I derive from
OpenThreads::Thread and have a OpenThreads::Mutex as a member of my new
thread class, then have a lock at the start of the parallell thread's main
loop waiting for a mutex to be released.
This works great for a week so until I test my app on windows and things
start to misbehave... Because on linux, with pthreads as the OpenThreads
backend, a lock is a lock is a lock and it doesn't matter from which thread
somebody tries to take a new lock, i.e. mutex is not recursive. While on
windows with CriticalSection (which is the Mutex backend on win32) the
thread that has the lock can relock it as many times as it likes, and from
what I read this by design and an important part of how Windows works.
Already ranting...will try to come to point.
All the above is ok, except that OpenThreads::Mutex works differently on
win32 and linux. Now - I have an implementation using Semaphores on windows
that makes Mutex work the same on win32 as with phtreads. My moves the
current behaviour to the OpenThreads::ReentrantMutex (which is currently
not implemented on win32). However - there is also some code in OpenThreads
that can only be enabled by manually defining USE_CRITICAL_SECTION that
leads me to believe that someone is already working on this or a similar
problem.
I'm getting more and more experienced in finding and fixing thread
synchronization problems (and want to avoid it as much as I can ;) , but
can't say I know everything about threads, so If you feel the urge to tell
me that i'm using the Mutex class in a bad way then please teach me bwana.
BUT the bottom line is still that the OpenThreads::Mutex class works
differerently on win32 and pthreads. We should fix it IMHO.
Avaiting your input or will post to osg-submissions (yes, it's a threat :)
/Mattias