> Tom Lane wrote
> I'm going to go ahead and make that change, since it doesn't
> seem likely
> to have any downside. It might be interesting to look into forcing
> proper alignment of the shared buffer headers as well.
Just catching up on your mails - all of that sounds good so far.
Everything mentioned so far talks about spinlocks in the general sense,
rather than with respect to particular locks.
The different lock types we have are held for different amounts of time
and are also subject to differing amounts of contention. I think it
would be useful to put in a capability to tune each class of lock
according to the possibility for delay and contention on it.
Long delay, low contention e.g. CheckpointLock => waiter should sleep
immediately
Medium delay, high contention e.g. WALWriteLock => waiter spins if at
head of queue, else sleeps immediately
Short delay, high contention e.g. BufMappingLock => waiter spins and
retries forever, cos the lock is coming soon
Short delay, low contention, sometimes long waits at end of VACUUM
e.g. FreeSpaceLock => waiter spins, then eventually sleeps
I'm not sure whether you'll agree with my characterisation of these
locks, but the main thing I'm trying to get across is that
once-size-fits-all isn't the optimal approach. I'm not saying either
that we end up with individual characterisations for each lock type, but
a few key ones could be catered for differently.
Best Regards, Simon Riggs
