> > I think what we ought to do is make ShmemAlloc act like palloc
> > (ie throw error not return NULL), and remove the duplicated error
> > checks. For the one caller that that would be bad for, we could
> > invent something like ShmemAllocNoError, or ShmemAllocExtended with
> > a no_error flag, or whatever other new API suits your fancy. But
> > as-is, it's just inviting more errors-of-omission like the large
> > number that already exist. I think people are conditioned by palloc
> > to expect such functions to throw error.
>
> The only reason why I did not propose that for ShmemAlloc is because
> of extensions potentially using this routine and having some special
> handling when it returns NULL. And changing it to behave like palloc
> would break such extensions.
I suggest to keep ShmemAlloc as is for backward compatibility and
introduce a new procedure ShmemAllocSafe. Also we could add a scary
comment (and also a warning log message?) that ShmemAlloc is deprecated
and possibly will be removed in later releases.
--
Best regards,
Aleksander Alekseev