On Thu, Aug 11, 2016 at 5:47 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> So, in my
> implementation, a condition variable wait loop looks like this:
>
> for (;;)
> {
> ConditionVariablePrepareToSleep(cv);
> if (condition for which we are waiting is satisfied)
> break;
> ConditionVariableSleep();
> }
> ConditionVariableCancelSleep();
I have what I think is a better idea. Let's get rid of
ConditionVariablePrepareToSleep(cv) and instead tell users of this
facility to write the loop this way:
for (;;)
{ if (condition for which we are waiting is satisfied) break; ConditionVariableSleep(cv);
}
ConditionVariableCancelSleep();
ConditionVariableSleep(cv) will check whether the current process is
already on the condition variable's waitlist. If so, it will sleep;
if not, it will add the process and return without sleeping.
It may seem odd that ConditionVariableSleep(cv) doesn't necessary
sleep, but this design has a significant advantage: we avoid
manipulating the wait-list altogether in the case where the condition
is already satisfied when we enter the loop. That's more like what we
already do in lwlock.c: we try to grab the lock first; if we can't, we
add ourselves to the wait-list and retry; if we then get the lock
after all we have to recheck whether we can get the lock and remove
ourselves from the wait-list if so. Of course, there is some cost: if
we do have to wait, we'll end up checking the condition twice before
actually going to sleep. However, it's probably smart to bet that
actually needing to sleep is fairly infrequent, just as in lwlock.c.
Thoughts?
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company