代码

https://github.com/gykimo/c_plusplus_optimize/tree/master/spinlock_vs_mutex

耗时

[case 1] mutex:

real    0m1.009s
user    0m0.744s
sys 0m0.865s

==========================
rm: spinlock_loop_nothing: No such file or directory
[case 2] spinlock_loop_nothing:

real    0m2.009s
user    0m3.959s
sys 0m0.019s

==========================
rm: spinlock_loop_sleep: No such file or directory
[case 3] spinlock_loop_sleep:

real    0m0.397s
user    0m0.521s
sys 0m0.261s

==========================
rm: spinlock_loop_nanosleep: No such file or directory
[case 4] spinlock_loop_nanosleep:

real    0m0.500s
user    0m0.766s
sys 0m0.207s

==========================
rm: spinlock_webrtc: No such file or directory
[case 5] spinlock_webrtc:

real    0m0.472s
user    0m0.853s
sys 0m0.064s

==========================
rm: mutex_heavy_task: No such file or directory
[case 6] mutex_heavy_task:
fail, count: 20000

real    0m3.135s
user    0m3.095s
sys 0m0.081s

==========================
rm: spinlock_webrtc_heavy_task: No such file or directory
[case 7] spinlock_webrtc_heavy_task:
fail, count: 20000

real    0m3.158s
user    0m5.706s
sys 0m0.445s

说明

当临界区耗时小时，如case1-case5，其中spinlock_webrtc user+sys耗时比mutex少很多，所以这种情况下自旋旋更合适；
当临界区耗时大时，如case6-case7，mutex的user+sys耗时比spinlock_webrtc少很多，所以这种情况下mutex更合适；

自旋锁实现不好，性能比mutex还低

如case1，Lock一直死循环判断lock_acquired是为0（空闲），反而性能很差，因为，lock_acquired是1期间，当线程一直满负载占用当前CPU，导致user耗时非常高。
代码如下：

    void Lock()
    {
        while (__sync_val_compare_and_swap(&lock_acquired, 0, 1))
        {
        }
    }

sleep版本性能好一些

    void Lock()
    {
        while (__sync_val_compare_and_swap(&lock_acquired, 0, 1))
        {
            sleep(0);
        }
    }

webrtc的性能最好

每次循环，调用sched_yield将当前线程挂载一会，这样可以不用一直占着CPU资源。

    void Lock()
    {
        while (__sync_val_compare_and_swap(&lock_acquired, 0, 1))
        {
            sched_yield();
        }
    }