Re: winealsa: Avoid deadlock in AudioClient_Stop.

[Andrew Eikum]

Hi Jörg,

Patch makes sense to me. I'll put together some patches for the other 
drivers, unless you beat me to it ;)

Thanks,
Andrew

On 06/17/2011 10:13 AM, [e-mail] wrote:
> Hi,
>
> my patch is not the final word about robust design.
>
> Consider what happens when
> DeleteTimerQueueTimer(g_timer_q, This->timer, INVALID_HANDLE_VALUE)
> cannot allocate its CompletionEvent.  The function will return with an
> (ignored) error and not stop the timers.  Do you see a check in
> wineoss or winecoreaudio?  The later Audio_Release will destroy
> critical sections and free memory while the timers are still running!
>
> A tiny allocation failure =>  big crash
>
> Deallocators should not fail when resources are scarce.
>
> A more robust design would see Audio_Start allocate the completion
> event that DeleteTimerQ requires and not start if it can't be sure to stop.
> Why require a completion event?  Because there's
> *no* other way of telling whether a callback might still be invoked.
> Even sleeping for more than the timer period is no guarantee rather
> than mere heuristics that works on a slightly loaded system.
>
> Back to my patch.  It's funny to think that it would mostly work
> without error checking: DeleteTimerQ(event) does something with
> event NULL or not, Wait(NULL) would immediately return and
> DeleteTimerQ always signals the event when successful.
>
> Regards,
>    Jörg Höhle