[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between
From: |
Paolo Bonzini |
Subject: |
Re: [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between disable and timer's cb |
Date: |
Tue, 30 Jul 2013 11:17:05 +0200 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:17.0) Gecko/20130625 Thunderbird/17.0.7 |
Il 30/07/2013 04:42, liu ping fan ha scritto:
> On Mon, Jul 29, 2013 at 7:21 PM, Paolo Bonzini <address@hidden> wrote:
>> Il 29/07/2013 10:10, liu ping fan ha scritto:
>>> On Mon, Jul 29, 2013 at 2:30 PM, Paolo Bonzini <address@hidden> wrote:
>>>> Il 29/07/2013 05:16, Liu Ping Fan ha scritto:
>>>>> After disabling the QemuClock, we should make sure that no QemuTimers
>>>>> are still in flight. To implement that, the caller of disabling will
>>>>> wait until the last user's exit.
>>>>>
>>>>> Note, the callers of qemu_clock_enable() should be sync by themselves,
>>>>> not protected by this patch.
>>>>>
>>>>> Signed-off-by: Liu Ping Fan <address@hidden>
>>>>
>>>> This is an interesting approach.
>>>>
>>>>> - if (!clock->enabled)
>>>>> - return;
>>>>> + atomic_inc(&clock->using);
>>>>> + if (unlikely(!clock->enabled)) {
>>>>> + goto exit;
>>>>> + }
>>>>
>>>> This can return directly, it doesn't need to increment and decrement
>>>> clock->using.
>>>>
>>> Here is a race condition like the following
>>
>> Ah, I see.
>>
>> Still this seems a bit backwards. Most of the time you will have no one
>> on the wait_using condvar, but you are almost always signaling the
>> condvar (should be broadcast BTW)...
>>
> I have tried to filter out the normal case by
> + qemu_mutex_lock(&clock->lock);
> + if (atomic_fetch_dec(&clock->using) == 1) { ---------> 1st place
> + if (unlikely(!clock->enabled)) { -------> 2nd place
> + qemu_cond_signal(&clock->wait_using);
> + }
> + }
> + qemu_mutex_unlock(&clock->lock);
>
> Could I do it better?
Hmm, do we even need clock->using at this point? For example:
qemu_clock_enable()
{
clock->enabled = enabled;
...
if (!enabled) {
/* If another thread is within qemu_run_timers,
* wait for it to finish.
*/
qemu_event_wait(&clock->callbacks_done_event);
}
}
qemu_run_timers()
{
qemu_event_reset(&clock->callbacks_done_event);
if (!clock->enabled) {
goto out;
}
...
out:
qemu_event_set(&eclock->callbacks_done_event);
}
In the fast path this only does two atomic operations (an OR for reset,
and XCHG for set).
Paolo
- [Qemu-devel] [RFC v2 0/5] arm AioContext with its own timer stuff, Liu Ping Fan, 2013/07/28
- [Qemu-devel] [RFC v2 1/5] timer: protect timers_state with lock, Liu Ping Fan, 2013/07/28
- [Qemu-devel] [RFC v2 2/5] timer: pick out timer list info from QemuClock, Liu Ping Fan, 2013/07/28
- [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between disable and timer's cb, Liu Ping Fan, 2013/07/28
- Re: [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between disable and timer's cb, Paolo Bonzini, 2013/07/29
- Re: [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between disable and timer's cb, liu ping fan, 2013/07/29
- Re: [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between disable and timer's cb, Paolo Bonzini, 2013/07/29
- Re: [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between disable and timer's cb, liu ping fan, 2013/07/29
- Re: [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between disable and timer's cb,
Paolo Bonzini <=
- Re: [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between disable and timer's cb, Alex Bligh, 2013/07/30
- Re: [Qemu-devel] [RFC v2 3/5] timer: make qemu_clock_enable sync between disable and timer's cb, Paolo Bonzini, 2013/07/30
[Qemu-devel] [RFC v2 4/5] timer: associate three timerlists with AioContext, Liu Ping Fan, 2013/07/28
[Qemu-devel] [RFC v2 5/5] timer: run timers on aio_poll, Liu Ping Fan, 2013/07/28
Re: [Qemu-devel] [RFC v2 0/5] arm AioContext with its own timer stuff, Stefan Hajnoczi, 2013/07/29