i have two threads, they run pretty fast, i'm using pthread_mutex_lock and pthread_mutex_unlock to access global (externed variables) data
the problem is that my application takes about 15-20% of CPU running on Ubuntu Linux,
the same code but with EnterCriticalSection and LeaveCriticalSection and running on Windows uses 1-2% of CPU
More important than CPU load is which version does more of real work. It could be that the 1-2% of load translates to less real work done. Measure that instead of the CPU load percentage.
That might be a good thing actually - less time spent in wait, more time crunching data.
CPU percentage calculations are very different on different OS-es. Try measuring your throughput - how many "work items" you are able to process in a unit of time.
One possible path for reducing lock contention (if this is indeed your problem) is to connect producer and consumer threads with a queue. Linking new item onto the queue tail is quick, same for un-linking off the queue head, - couple of pointer operations. STL even has a bunch of containers you can use (std::deque, std::queue, std::list). You would have to provide your own locking though. Or look into Intel Threading Building Blocks.
I can't quite tell what your question is, but let's say it's "How do I make the Linux version faster?".
First, are you sure that on Linux you enabled optimization?
Assuming yes, are both programs doing the same amount of "work"?
If that's the case, then you need to profile - that will show you directly where the CPU cycles are being used and should enable you to optimize your algorithm/code.