ARTICLE AD BOX
I'm taking a course about concurrency and lock-free programming, we touched on RCU (Read-Copy-Update) as a way to do memory reclamation. Assume we know all threads involved have an ID from 0 to THREADS-1, and we can get the current thread's ID by calling get_id. The following toy implementation of RCU was provided:
// shared data std::atomic<int> current_phase{}; std::array<std::atomic<int>, THREADS> rcu_phase{}; std::array<std::atomic<int>, THREADS> rcu_rcs{}; void rcu_read_lock() { int tmp = current_phase.load(std::memory_order_acquire); rcu_phase[get_id()].store(tmp, std::memory_order_release); rcu_rcs[get_id()].store(1, std::memory_order_release); std::atomic_thread_fence(std::memory_order_seq_cst); } void rcu_read_unlock() { rcu_rcs[get_id()].store(0, std::memory_order_release); } void rcu_synchronize() { std::atomic_thread_fence(std::memory_order_seq_cst); current_phase.store(1, std::memory_order_release); for (size_t i = 0; i < THREADS; i++) { do { int a = rcu_rcs[i].load(std::memory_order_acquire); int b = rcu_phase[i].load(std::memory_order_acquire); } while(a == 1 && b == 0); } for (size_t i = 0; i < THREADS; i++) { do { int a = rcu_rcs[i].load(std::memory_order_acquire); int b = rcu_phase[i].load(std::memory_order_acquire); } while(a == 1 && b == 1); } }There are at least three things I don't undestand about this:
What is the need of having rcu_phase? Isn't having rcu_rcs sufficient as that is what signals if a reader thread is actually reading the critical section?
If having the phases makes sense, why does rcu_synchronize need switch between twice? Couldn't it simply look what the current phase is and flip it, and then wait for the threads from the previous phase?
What is the purpose of those two sequentially consistent fences?
Thank you for help! Online resources about this are either very shy of going into the implementation details and mostly explain how to use the API or they refer to more complex/different implementations.
