Atomic swap in GNU C++

问题:

I want to verify that my understanding is correct. This kind of thing is tricky so I'm almost sure I am missing something. I have a program consisting of a real-time thread and a non-real-time thread. I want the non-RT thread to be able to swap a pointer to memory that is used by the RT thread.

From the docs, my understanding is that this can be accomplished in g++ with:

// global Data *rt_data;  Data *swap_data(Data *new_data) { #ifdef __GNUC__     // Atomic pointer swap.     Data *old_d = __sync_lock_test_and_set(&rt_data, new_data); #else     // Non-atomic, cross your fingers.                                               Data *old_d = rt_data;     rt_data = new_data; #endif     return old_d; } 

This is the only place in the program (other than initial setup) where rt_data is modified. When rt_data is used in the real-time context, it is copied to a local pointer. For old_d, later on when it is sure that the old memory is not used, it will be freed in the non-RT thread. Is this correct? Do I need volatile anywhere? Are there other synchronization primitives I should be calling?

By the way I am doing this in C++, although I'm interested in whether the answer differs for C.

Thanks ahead of time.

回答1:

Generally don't use volatile when writing concurrent code in C/C++. The semantics of volatile are so close to what you want that it is tempting but in the end volatile is not enough. Unfortunately Java/C# volatile != C/C++ volatile. Herb Sutter has a great article explaining the confusing mess.

What you really want is a memory fence. __sync_lock_test_and_set provides the fencing for you.

You will also need a memory fence when you copy (load) the rt_data pointer to your local copy.

Lock free programming is tricky. If you're willing to use Gcc's c++0x extensions, it's a bit easier:

#include   std::atomic rt_data;  Data* swap_data( Data* new_data ) {    Data* old_data = rt_data.exchange(new_data);    assert( old_data != new_data );    return old_data; }  void use_data( ) {    Data* local = rt_data.load();    /* ... */ } 

回答2:

Update: This answer is not correct, as I am missing the fact that volatile guarantees that accesses to volatile variables are not reordered, but provides no such guarantees with respect to other non-volatile accesses and manipulations. A memory fence does provide such guarantees, and is necessary for this application. My original answer is below, but do not act on it. See this answer for a good explanation in the hole in my understanding that led to the following incorrect response.

Original answer:

Yes, you need volatile on your rt_data declaration; any time a variable can be modified outside the flow of control of a thread accessing it, it should be declared volatile. While you may be able to get away without volatile since you're copying to a local pointer, volatile at least helps with documentation and also inhibits some compiler optimizations that can cause problems. Consider the following example, adopted from DDJ:

volatile int a; int b; a = 1; b = a; 

If it is possible for a to have its value changed between a=1 and b=a, then a should be declared volatile (unless, of course, assigning an out-of-date value to b is acceptable). Multithreading, particularly with atomic primitives, constitutes such a situation. The situation is also triggered with variables modified by signal handlers and by variables mapped to odd memory locations (e.g. hardware I/O registers). See also this question.

Otherwise, it looks fine to me.

In C, I would probably use the atomic primitives provided by GLib for this. They'll use an atomic operation where available and fall back to a slow-but-correct mutex-based implementation if the atomic operations are not available. Boost may provide something similar for C++.

文章来源: Atomic swap in GNU C++