Re: [PATCHES] Win32 CHECK_FOR_INTERRUPTS() performance

On Mon, 24 Oct 2005, Tom Lane wrote:

> Qingqing Zhou <zhouqq(at)cs(dot)toronto(dot)edu> writes:
> > I tried to persuade myself that removing all WaitForSingleObjectEx() is
> > safe ... the thing is we will false alarm EINTR as Magnus said (details to
> > repeat it are list below in case).
>
> Just to repeat myself: there were false alarms before. The interleaving
> you describe could equally well happen if a new signal is sent just
> after the old code executes WaitForSingleObjectEx and sees that a
> previous signal is waiting for it. Both old and new signals can be
> cleared by the recipient before the second signal sender gets as far as
> setting the event.
>

Oh, yeah. Just write the detailed case down for the sake of memory:

-- For previous code -- false alarm case --
1. I am killing you signal A:
enter_crit;
set signal bit;
leave_crit;

2. He *has killed* you signal B:

3. You CHECK_FOR_INTERRUPTS():
enter_crit;
sig(A);
sig(B);
ResetEvent();
leave_crit;

4. I finish my killing:
SetEvent();

Now the event is signaled but the signal is handled already.

Regards,
Qingqing