Hi,
On 2020-06-09 19:24:15 -0400, Tom Lane wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
> > On Tue, Jun 9, 2020 at 1:59 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> >> When I went through the existing spinlock stanzas, the only thing that
> >> really made me acutely uncomfortable was the chunk in pg_stat_statement's
> >> pgss_store(), lines 1386..1438 in HEAD.
>
> > I mean, what would be wrong with having an LWLock per pgss entry?
+1
> Hmm, maybe nothing. I'm accustomed to thinking of them as being
> significantly more expensive than spinlocks, but maybe we've narrowed
> the gap enough that that's not such a problem.
They do add a few cycles (IIRC ~30 or so, last time I measured a
specific scenario) of latency to acquisition, but it's not a large
amount. The only case where acquisition is noticably slower, in my
experiments, is when there's "just the right amount" of
contention. There spinning instead of entering the kernel can be good.
I've mused about adding a small amount of spinning to lwlock acquisition
before. But so far working on reducing contention seemed the better
route.
Funnily enough lwlock *release*, even when there are no waiters, has a
somewhat noticable performance difference on x86 (and other TSO
platforms) compared to spinlock release. For spinlock release we can
just use a plain write and a compiler barrier, whereas lwlock release
needs to use an atomic operation.
I think that's hard, but not impossible, to avoid for an userspace
reader-writer lock.
It would be a nice experiment to make spinlocks a legacy wrapper around
rwlocks. I think if we added 2-3 optimizations (optimize for
exclusive-only locks, short amount of spinning, possibly inline
functions for "fast path" acquisitions/release) that'd be better for
nearly all situations. And in the situations where it's not, the loss
would be pretty darn small.
Greetings,
Andres Freund