bt2_locks.h

#ifndef __MCSLOCK_H__
#define __MCSLOCK_H__

#include <atomic>
#include <sched.h>

/*
 * Based on TBB's atomic_backoff: https://github.com/oneapi-src/oneTBB/blob/60b7d0a78f8910976678ba63a19fdaee22c0ef65/include/tbb/tbb_machine.h
 */
class cpu_backoff {
public:
	cpu_backoff(): count(1) {}
	inline void pause() {
		if (count <= LOOPS_BEFORE_YIELD) {
			for (int32_t i = 0; i < count; i++) {
#ifdef __aarch64__
				__asm__ __volatile__("yield" ::: "memory");
#elif __ppc__
				__asm__ __volatile__("or 27,27,27" ::: "memory");
#elif __x86_64__
				__asm__ __volatile__("pause;");
#else
				// do nothing
#endif
			}
			count *= 2;
		} else {
			sched_yield();
		}
	}
private:
	static const int32_t LOOPS_BEFORE_YIELD = 16;
	int32_t count;
};

class spin_lock {
	std::atomic_flag flag;
public:
	spin_lock() : flag(false) {}

	void lock();
	void unlock();
};

class mcs_lock {
public:
	mcs_lock(): q(nullptr) {}
        struct mcs_node {
		std::atomic<mcs_node *> next;
		std::atomic_bool unlocked;
        };

	void lock(mcs_node &node);
	void unlock(mcs_node &node);
	typedef std::atomic<mcs_node *> mcs_node_ptr;
private:
	void spin_while_eq(const volatile mcs_node_ptr& value, const volatile mcs_node *expected) {
		cpu_backoff backoff;
		while (value.load(std::memory_order_acquire) == expected)
			backoff.pause();
	}

	void spin_while_eq(const volatile std::atomic_bool& value, const volatile bool expected) {
	 	cpu_backoff backoff;
	 	while (value.load(std::memory_order_acquire) == expected)
	 		backoff.pause();
	}
        std::atomic<mcs_node *> q;
};

#endif // __MCS_LOCK_H__