Re: [Qemu-devel] [PATCH 07/12] qemu-thread: add QemuSemaphore

[Paolo Bonzini <pbonzini@redhat.com>]

Il 16/07/2012 16:09, Jan Kiszka ha scritto:
> On 2012-07-16 16:03, Paolo Bonzini wrote:
>> Il 16/07/2012 15:53, Jan Kiszka ha scritto:
>>>>>
>>>>> qemu_cond_wait only uses WaitForSingleObject with INFINITE timeout, and
>>>>> the algorithm relies on that.
>>> I see. But this doesn't look complex awfully. Just move the waker
>>> signaling from within cond_wait under the mutex as well, maybe add
>>> another flag if there is really someone waiting, and that's it. The
>>> costs should be hard to measure, even in line of code.
>>
>> There is still a race after WaitForSingleObject times out.  You need to
>> catch the mutex before decreasing the number of waiters, and during that
>> window somebody can broadcast the condition variable.
> 
> ...and that's why you check what needs to be done to handle this race
> after grabbing the mutex. IOW, replicate the state information that the
> Windows semaphore contains into the emulated condition variable object.

It is already there (cv->waiters), but it is accessed atomically.  To do
what you suggest I would need to add a mutex.

>> I'm not saying it's impossible, just that it's hard and I dislike
>> qemu_cond_timedwait as much as you dislike semaphores. :)
> 
> As you cannot eliminate our use cases for condition variables, let's
> focus on the existing synchronization patterns instead of introducing
> new ones + new mechanisms.

I'm not introducing <fancy new primitive of the day>... semaphores have
the advantage of having better support under Windows (at least older
versions; Microsoft saw the light starting with Vista---yes I know the
oxymoron---and added condvars).

Paolo