Tom Lane wrote:
> Um. I don't like that; it assumes not only that ExecutorEnd is the only
> kind of callback needed, but also that there is at most one function
> per ExprContext that needs a shutdown callback. Neither of these
> assumptions hold water IMO.
>
> The design I had in mind was more like this: add to ExprContext a list
> header field pointing to a list of structs along the lines of
>
> struct exprcontext_callback {
> struct exprcontext_callback *next;
> void (*function) (Datum);
> Datum arg;
> }
>
> and then call each specified function with given argument during
> FreeExprContext. Probably ought to be careful to do that in reverse
> order of registration. We'd also need to invent a RescanExprContext
> operation to call the callbacks during a Rescan. The use of Datum
> (and not, say, void *) as PG's standard callback arg type was settled on
> some time ago --- originally for on_proc_exit IIRC --- and seems to have
> worked well enough.
Well, I guess I set my sights too low ;-) This is a very nice design.
I have the shutdown callback working now, and will send a new patch in a
few minutes. I have not started RescanExprContext() yet, but will do it
when I address rescans in general.
> What's really more relevant here is that during the hypothetical new
> RescanExprContext function, we ought to go around and clear any fcaches
> in the context that have setArgsValid = true, so that they will be
> restarted afresh during the next scan of the plan. (The fact that that
> doesn't happen now is another shortcoming of the existing set-functions-
> in-expressions code.) So this suggests making a callback function type
> specifically to do that, and registering every fcache that is executing
> a set function in the callback list...
I also added FunctionCachePtr_callback struct and a member to
ExprContext. I have not yet created the registration or shutdown
functions, but again, I'll work on them as part of the rescan work.
I still have a couple of issues related to VIEWs that I need to figure
out, then I'll start the rescan work.
Thanks for the review and help!
Joe
