According to cppreference, std::atomic::notify_one() will notify at least one thread that is waiting on said atomic. This means that according to the standard it could unblock more than one thread. This is in contrast to std::condition_variable::notify_one(), which specifies that it will unblock (no more than) one thread.
Where does this difference come from? Does this not use the same underlying mechanism? As far as implementations of the standard library go, do all of the prevalent ones have a chance of actually unblocking multiple with this call, or are there some that always unblock exactly one?
CodePudding user response:
As far as I can tell, they can use the same mechanism, but don't have to. The implementation I typically use is libstdc and its atomic::notify_one() is definitely different.
The condition_variable is fairly opaque, implemented as __condvar in the bits/std_mutex.h header. Its notify_one() just calls a pthread wrapper:
void
notify_one() noexcept
{
int __e __attribute__((__unused__)) = __gthread_cond_signal(&_M_cond);
__glibcxx_assert(__e == 0);
}
whereas notify_all() instead calls __gthread_cond_broadcast.
This contrasts with the atomics, which are much more involved, but end up using a sequence of functions for notifying threads, starting with __atomic_notify_address:
template<typename _Tp>
void
__atomic_notify_address(const _Tp* __addr, bool __all) noexcept
{
__detail::__bare_wait __w(__addr);
__w._M_notify(__all);
}
void
_M_notify(bool __all, bool __bare = false)
{
if (_M_laundered())
{
__atomic_fetch_add(_M_addr, 1, __ATOMIC_SEQ_CST);
__all = true;
}
_M_w._M_notify(_M_addr, __all, __bare);
}
__waiter_pool_base::_M_notify:
void
_M_notify(const __platform_wait_t* __addr, bool __all, bool __bare) noexcept
{
if (!(__bare || _M_waiting()))
return;
#ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
__platform_notify(__addr, __all);
#else
if (__all)
_M_cv.notify_all();
else
_M_cv.notify_one();
#endif
}
So if there is no _GLIBCXX_HAVE_PLATFORM_WAIT, libstdc will simply use a condition variable to notify either one or all threads. Otherwise, it's implemented using a futex, and may also wake (up to?) INT_MAX threads, see __platform_notify:
template<typename _Tp>
void
__platform_notify(const _Tp* __addr, bool __all) noexcept
{
syscall (SYS_futex, static_cast<const void*>(__addr),
static_cast<int>(__futex_wait_flags::__wake_private),
__all ? INT_MAX : 1);
}
CodePudding user response:
Both std::atomic::wait
and std::condition_variable::wait
are allowed to unblock spuriously at any time, including at the specific time that notify_one
is called.
So in terms of the specification, although it does indeed seem that the standard uses different wording for the two ("at least one" in [atomics.types.operation]/32 but just "one" in [thread.condition.condvar]/5), I don't think there is any normative difference here. The implementation is free to unblock any number of threads for both waiting operations when notify_one
is called, as long as it is at least one (if any are waiting). This is true for std::atomic::notify_one
as well as std::condition_variable::notify_one
(i.e. "(no more than)" is wrong).
In an earlier revision of the proposal [P1135R4] for the std::atomic
waiting operations it still said "one" instead of "at least one" in the proposed wording.
That was changed with revision [P1135R5], according to which wording improvements were made partly in response to feedback from LWG teleconferences. I don't think any documentation of these teleconferences is publicly available, so I can't tell whether there was a specific intent in this wording change.
Maybe it is supposed to avoid misunderstandings since the spurious unblocking is mentioned in a different paragraph or maybe it is meant to convey that there are implementations which do unblock multiple/all waiting operations that are specifically intended to be supported as such. See also comments under the question and this answer.