“OTP 源码编译”的版本间的差异

2017年1月7日 (六) 12:45的最后版本

[编辑] OTP 源码阅读杂记

[编辑] 编译

git clone https://github.com/erlang/otp.git
cd otp
export ERL_TOP=`pwd`
./otp_build  setup
brew install fop openssl
./configure --help #查看编译选项
./configure --with-ssl=/usr/local/opt/openssl
make

[编辑] erl_process.c

• 优先运行 port 直到完成。

	/*
	 * Find a new port to run.
	 */

	if (RUNQ_READ_LEN(&rq->ports.info.len)) {
	    int have_outstanding_io;
	    have_outstanding_io = erts_port_task_execute(rq, &esdp->current_port);
	    if ((!erts_eager_check_io
		 && have_outstanding_io
		 && fcalls > 2*input_reductions)
		|| rq->halt_in_progress) {
		/*
		 * If we have performed more than 2*INPUT_REDUCTIONS since
		 * last call to erl_sys_schedule() and we still haven't
		 * handled all I/O tasks we stop running processes and
		 * focus completely on ports.
		 *
		 * One could argue that this is a strange behavior. The
		 * reason for doing it this way is that it is similar
		 * to the behavior before port tasks were introduced.
		 * We don't want to change the behavior too much, at
		 * least not at the time of writing. This behavior
		 * might change in the future.
		 *
		 * /rickard
		 */
		goto check_activities_to_run;
	    }
	}

• 垃圾回收也消耗 reductions:

int cost = scheduler_gc_proc(p, reds);
		calls += cost;
		reds -= cost;

[编辑] erl_lock_check.c 和锁

• 确保 BEAM 的加锁顺序是一致的，防止死锁。

 * Description: A lock checker that checks that each thread acquires
 *              locks according to a predefined global lock order. The
 *              global lock order is used to prevent deadlocks. If the
 *              lock order is violated, an error message is printed
 *              and the emulator aborts. The lock checker is only
 *              intended to be enabled when debugging.

• include/internal/ 下面是各平台的锁的实现，比如 spinlock, rwlock， x86 平台在 i386 目录下，比如 spinlock.h，使用 xchgb 指令汇编实现

static ETHR_INLINE int
ethr_native_spin_trylock(ethr_native_spinlock_t *lock)
{
    char tmp = 1;
    __asm__ __volatile__(
	"xchgb %b0, %1"
	: "=q"(tmp), "=m"(lock->lock)
	: "0"(tmp) : "memory");
    return tmp == 0;
}
static ETHR_INLINE void
ethr_native_spin_lock(ethr_native_spinlock_t *lock)
{
    for(;;) {
	if (__builtin_expect(ethr_native_spin_trylock(lock) != 0, 1))
	    break;
	do {
	    __asm__ __volatile__("rep;nop" : "=m"(lock->lock) : : "memory");
	} while (ethr_native_spin_is_locked(lock));
    }
}

如果没有特定平台的实现，那就使用 pthread_spin_lock， native 的实现没有采用 backoff 或者 queue 技术来优化。但是 "rep;nop" 等价于 pause 指令，这是为了 spinlock 性能优化引入的。参考 http://stackoverflow.com/questions/7086220/what-does-rep-nop-mean-in-x86-assembly