On Sun, Jun 6, 2010 at 4:53 PM, Avi Kivity <a...@redhat.com> wrote:
> On 06/06/2010 05:48 PM, Corentin Chary wrote:
>>
>> On Sun, Jun 6, 2010 at 4:11 PM, Avi Kivity<a...@redhat.com> wrote:
>>
>>>
>>> On 06/04/2010 04:20 PM, Corentin Chary wrote:
>>>
>>>>
>>>> + if (vnc_trylock_display(vd)) {
>>>> + vd->timer_interval = VNC_REFRESH_INTERVAL_BASE;
>>>> + qemu_mod_timer(vd->timer, qemu_get_clock(rt_clock) +
>>>> + vd->timer_interval);
>>>> + return;
>>>> + }
>>>> +
>>>> has_dirty = vnc_refresh_server_surface(vd);
>>>> + vnc_unlock_display(vd);
>>>>
>>>>
>>>
>>> This could delay the update by quite a bit, no?
>>>
>>
>> Yep, but it's far better than waiting the lock because it doesn't slow
>> down the main thread.
>> I played big buck bunny trailler (33sec) in mplayer and tight encoding:
>> - ~40 sec with the non-threaded server
>> - ~37 sec with a lock
>> - ~33 sec with a try_lock
>>
>
> Definitely, blocking the main thread is a no-no.
>
>>> A more elaborate approach would be to enqueue the refresh job into the
>>> queue. May need the iothread enabled so we have qemu_mutex.
>>>
>>
>> Maybe, but I'd like to wait the generic async work subsystem before
>> adding different kind of jobs to the queue. And it's already a big
>> improvment over the current code :).
>>
>
> Hm, ok.
>
>>> btw, I could not find other uses of vd->mutex, shouldn't it protect
>>> against
>>> the work thread?
>>>
>>
>> Check vnc-jobs.c, there is a qemu_mutex_lock(vs->vd->mutex);
>>
>>
>
> Shouldn't it use vnc_lock_display()? That's why I missed it.
I didn't use vnc_lock_display because I didn't want to export it first.
Maybe I should also use vnc_lock_output() in vnc-jobs.c ...
--
Corentin Chary
http://xf.iksaif.net
