Re: Performance degradation in commit ac1d794

Hello, I am one who wants waiting on many sockets at once.

At Thu, 14 Jan 2016 18:55:51 +0100, Andres Freund <andres(at)anarazel(dot)de> wrote in <20160114175551(dot)GM10941(at)awork2(dot)anarazel(dot)de>
> On 2016-01-14 18:14:21 +0100, Andres Freund wrote:
> > On 2016-01-14 12:07:23 -0500, Robert Haas wrote:
> > > > Do we want to provide a backward compatible API for all this? I'm fine
> > > > either way.
> > >
> > > How would that work?
> >
> > I'm thinking of something like;
> >
> > int WaitOnLatchSet(LatchEventSet *set, int wakeEvents, long timeout);
> >
> > int
> > WaitLatchOrSocket(volatile Latch *latch, int wakeEvents, pgsocket sock,long timeout)
> > {
> > LatchEventSet set;
> >
> > LatchEventSetInit(&set, latch);
> >
> > if (sock != PGINVALID_SOCKET)
> > LatchEventSetAddSock(&set, sock);
> >
> > return WaitOnLatchSet(set, wakeEvents, timeout);
> > }
> >
> > I think we'll need to continue having wakeEvents and timeout parameters
> > for WaitOnLatchSet, we quite frequently want to wait socket
> > readability/writability, not wait on the socket, or have/not have
> > timeouts.
>
> Hm. If we really want to support multiple sockets at some point the
> above WaitOnLatchSet signature isn't going to fly, because it won't
> support figuring out which fd the event this triggered on.
>
> So it seems we'd need to return something like
> struct LatchEvent
> {
> enum LatchEventType {WL_LATCH_EVENT;WL_TIMEOUT;WL_SOCKET_EVENT; WL_POSTMASTER_EVENT;...} event_type;
> int mask;
> pgsocket event_sock;
> };
>
> that'd also allow to extend this to return multiple events if we want
> that at some point. Alternatively we could add a pgsocket* argument, but
> that doesn't really seem much better.
>
> Not super happy about the above proposal.

How about allowing registration of a callback for every waiting
socket. The signature of the callback function would be like

enum LATCH_CALLBACK_STATE
LatchWaitCallback(pgsocket event_sock,
enum LatchEventType, int mask?, void *bogus);

It can return, for instance, LCB_CONTINUE, LCB_BREAK or
LCB_IMMEDBREAK, and if any one of them returns LCB_BREAK, it will
break after, maybe, calling all callbacks for fired events.

We could have predefined callbacks for every event which does
only setting a corresponding flag and returns LCB_BREAK.

/* Waiting set has been constructed so far */
if (!WaitOnLatchSet(&set(?))
errorr()

if (is_sock_readable[sockid]) {} /** is... would be global /
if (is_sock_writable[sockid]) {} /** is... would be global /
/* Any other types of trigger would processes elsewhere */

Although it might be slow if we have an enormous number of
sockets fired at once, I suppose it returns only for a few
sockets on most cases.

I don't see the big picture of the whole latch/signalling
mechanism with this, but callbacks might be usable for signalling
on many sockets.

regards,

--
Kyotaro Horiguchi
NTT Open Source Software Center