Bug in XPATH() produces invalid XML values and probably un-restorable dumps

Florian G. Pflug <fgp@phlo.org>

From: Florian Pflug <fgp@phlo.org>
To: PG Hackers <pgsql-hackers@postgresql.org>
Date: 2011-05-31T18:50:23Z
Lists: pgsql-hackers

Attachments

Hi

While trying to figure out sensible semantics for XPATH() and scalar-value returning XPath expressions, I've stumbled upon a bug in XPATH() that allows invalid XML values to be produced. This is a serious problem because should such invalid values get inserted into an XML column, an un-restorable dump ensues.

Here's an example (REL9_0_STABLE as of a few days ago)

template1=# SELECT (XPATH('/*/text()', '<root>&lt;</root>'))[1];
 xpath 
-------
 <

Since XPATH() returns XML[], this value has type XML, but clearly isn't well-formed. And behold, casting to TEXT and back to XML complains loudly.

template1=# SELECT (XPATH('/*/text()', '<root>&lt;</root>'))[1]::TEXT::XML;
ERROR:  invalid XML content
DETAIL:  Entity: line 1: parser error : StartTag: invalid element name
	<
	 ^

The culprit is xml_xmlnodetoxmltype() in backend/utils/adt/xml.c. For non-element nodes, it returns the result of xmlXPathCastNodeToString() verbatim, even though that function doesn't reverse the entity replacement that was done during parsing. Adding a call to escape_xml() 
for non-element nodes fixes the problem

template1=# SELECT (XPATH('/*/text()', '<root>&lt;</root>'))[1];
 xpath 
-------
 &lt;

Patch is attached.

best regards,
Florian Pflug