On Sat, Oct 21, 2023 at 2:21 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2023-10-20 18:16:04 -0400, Tom Lane wrote:
> >> That must be coming from dsm_impl_pin_segment or dsm_impl_unpin_segment;
> >> noplace else has the same error message spelling.
>
> > I guess we might now parallelize queries that we didn't in 11 and thus it's
> > more likely that we end up needing the typmod registry?
>
> Maybe. Disabling parallel queries should be an easy stopgap workaround,
> in any case.
Yeah.
Some initial observations about the error: it's
dsm_impl_pin_segment(), not dsm_impl_unpin_segment() (kinda strange
that we have the same error message for the latter given that it's
trying to close it, but anyway here we can see the line number in the
OP's report). Both callers of dsm_pin_segment() are immediately after
dsm_create(), so we're dealing with an inability to pin a freshly
allocated memory object. The way we pin in Windows is that we
duplicate the handle into the postmaster for safe-keeping, to keep its
refcount (the kernel's refcount, not ours) from reaching zero, and
then we unpin by closing that dup'd handle in the postmaster. Perhaps
there are two handles that could be b0rked there, triggering an error
message like that: the postmaster process handle, and the memory
object handle.