On Jul26, 2011, at 15:52 , Tom Lane wrote:
> Florian Pflug <fgp@phlo.org> writes:
>> The only fix I can come up with is to explicitly test whether or not we're
>> able to restore the structured error context in pg_xml_init_library().
>
> Yeah, I think you are right.
>
>> The downside of this is that a libxml version upgrade might break postgres,
>
> Such an upgrade would break Postgres anyway --- better we are able to
> detect and report it clearly than that we fail in bizarre ways.
On further reflection, instead of checking whether we can restore the error
handler in pg_xml_init_library(), we could simply upgrade the elog(WARNING)
in pg_xml_done() to ereport(ERROR), and include a hint that explains the
probably cause.
The upside being that we only fail when we actually need to restore the
error handler. Since there's one caller (parse_xml_decl) who calls
pg_xml_init_library() but not pg_xml_init()/pg_xml_done(), the difference
isn't only academic. At least XML would output will continue to work
work after libxml is upgraded from < 2.7.4 to >= 2.7.4.
If nobody objects, I'll post a patch to that effect later today.
BTW, come to think of it, I believe the elog(ERROR) that you put into
xml_errorHandler should to be upgraded to elog(FATAL). Once we longjmp()
out of libxml, it doesn't seem safe to re-enter it, so making the backend
exit seems to be the only safe thing to do.
best regards,
Florian Pflug