st 19. 2. 2025 v 19:05 odesílatel Andres Freund <andres@anarazel.de> napsal:
Hi,
On 2025-02-19 01:48:53 -0500, Tom Lane wrote: > Pavel Stehule <pavel.stehule@gmail.com> writes: > > I investigated the crashes in "xmlnamespaces to xmlelement" patch and it > > looks like there was a badly used makeString function. The argument should > > not be null, elsewhere serialization to string fails - and deserialization > > doesn't support this case. > > I propose to add an assert there like (make check-world passed) > > Hmmm ... while I don't necessarily object to this patch, we have a lot > of makeFoo() functions that build nodes, and hardly any of them have > asserts like this one.
I also suspect that adding is-not-NULL asserts isn't that helpful on its own, because you still need to reach that function with it set to NULL. We probably should use pg_attribute_nonnull() much more widely, so that compilers and static analyzers can help.
Any mechanism that can help is welcome. Unfortunately, I miss knowledge about C static analyzers.
This issue can be pretty messy for beginners - who try to write their own first patches. The error is raised
far to the point where this problem was created, and it needs a special test or special configuration setting.
Probably this issue is less often than before, because out functions are generated automatically, but there
can be bad design. Another possibility is implementing correct support for NULL there, but probably it needs
format change, and it can be an issue for pg_upgrade.
Regards
Pavel
> Why makeString() in particular? Is the fault on the serialization side, > instead? If there's a general expectation that a String node's value isn't > null, how come the original patch worked at all?
It's worth noting that the CI task just failed on freebsd, which builds with:
