W3C home > Mailing lists > Public > public-xmlsec@w3.org > August 2010

Streamable XPath Profile Review

From: Meiko Jensen <Meiko.Jensen@ruhr-uni-bochum.de>
Date: 9 Aug 2010 17:09:57 +0200
Message-ID: <4C601A45.4000401@ruhr-uni-bochum.de>
To: "XMLSec WG Public List" <public-xmlsec@w3.org>

I took a closer look at the latest version of the Streamable XPath
Profile, and came across a set of issues I considered important to be

1. The document by now is structured as a kind of "limitation of XPath
1.0". This means it is very hard to read (or understand) it without
having the XPath 1.0 spec next to it. I'm not sure this is the ideal way
to shape the spec, though being a "profile" for XPath.

2. The document relies on XPath Version 1.0, though XPath 2.0 is out. I
think this is no big issue since XPath 2.0 (for some reason) does not
supersede XPath 1.0, but we might run into questions regarding this from
other W3C groups?

3. To dive into the contents, the specified XPath subset definitely is
streamable. All streaming-critical components (such as backward axes and
complex XPaths in predicates) have been excluded.

4. There are some special properties coming up with this profile. I'll
try to resume briefly:

4a. Allowing the descendant axis implies that it is possible to have
XPaths resulting in more than one context node at each step. For
instance, evaluating "//A//B" in the XML document
"<A><B><A><A><B/></A></A></B></A>" has up to three different context
nodes during its evaluation, and ends up with two nodes being part of
the result, one of them being result of three different evaluation
paths. We've been talking about this before. It is valid to have such
selections, it just makes evaluation become more complex (potential for
Denial of Service here!!), since you have to maintain more than one
context node per streaming event.

4b. Similar to 4a, having "|" allowed implies possibility of more than
one context node.

4c. the profile explicitly excludes the use of the "text()" and "node()"
operators in order not to run into complex text node collection issues,
which are very difficult in streaming evaluators. However, XPath 1.0 has
some loopholes that still run into selecting texts. For instance, the
XPath 1.0 function of "string()" without explicit argument implies to
convert the context node to string, which means concatenating all its
text node descendants. Hence, we have implicit text() selection here. In
the same line, I'm not yet convinced whether other conversion functions
of XPath 1.0 (any "node-set to X conversion") might also result in
implicit access to data that is not available in the streaming model.
This would require an in-depth compatibility analysis between the
Streamable XPath profile and XPath 1.0.

4d. Excluding the "node()" operator but allowing the "." and "//"
operators is a contradicition per se, since XPath 1.0 says "." is an
abbreviation of "self::node()" and "//" means
"/descendant-or-self::node()/". However, as long as we explicitly
explain the valid and invalid use cases of "node()" we should be fine.

4e. The "following" and "following-sibling" axes in fact are streamable,
yet I don't see a good reason to keep them in this profile since they
would make evaluation get way more complex again.

4f. The given grammar explicitly disallows predicates in non-final XPath
steps. This contradicts with example no. 5 that says
"/book/chapter[2]/title[1]" would be allowed. In fact, "chapter[2]" is
an abbreviation of "chapter[position()=2]", hence being a regular
predicate. In this case I'd strongly recommend to allow position
indicators to be used in non-final XPath steps. This is one of the major
requirements for fending signature wrapping attacks: the XPath
expression must pin down the signed element's location precisely, hence
would definitely need position indicators at every step. I'd even go so
far as to recommend having position indicators being REQUIRED for every

4g. The specified subset of XPath in fact is very close to the FastXPath
subset we've been proposing in 2009
One of the interesting differences, though, is that FastXPath also
prohibits the use of the "or" operator within predicates. The rationale
behind that was to stick to the "policy of reduction": during
evaluation, every test in every step of the XPath either keeps a
(single) context node or switches to the empty result set (and
immediately cancels further evaluation). The "or" operator violates this
policy: if the first condition evaluates to "false" (=> result-set gets
empty), it is possible that the second condition might become "true",
putting the context node back in to the result-set. Again, this causes
additional parsing complexity (Denial of Service via huge amount of "or"
conditions). However, this was a design decision for FastXPath that I'd
recommend to consider for the Streamable XPath Profile, but I do not
insist in its adoption.

5. Some minor issues:

5a. The document needs some typo/spell checking.

5b. in "1. Introduction" list item "2.b." says that "/book/chapter[3]"
is not expressable in XMLDSig's XPath Filter.
"ancestor-or-self::chapter[ . = /book/chapter[3] ]" should do, yet I
agree with the intention behind this sentence that it is way too complex.

5c. in "1. Introduction" list item "3.b." says that "ID based references
are streamable". If you consider the IDed element to occur before the
XML signature within the XML document, this does not hold. I'd soften
this sentence to something like "ID references usually are easier to
process than XPath Filter 2 Transforms in a streaming-based parsing

5e. in "4. Definition of the Profile" the notes to "[1] LocationPath
::=" say that relative location paths are not allowed, claiming these to
start from the "<Signature>" element. I'd suggest to remove this comment
for now, since it brings up false assumptions just to mark them invalid
immediately. However, since I still consider such a referencing scheme
(with relative XPath starting at a context node different from the
document root) useful, I'd suggest discussing this issue separately,
then decide how to deal with it in the XPath profile.

5f. in "6. Algorithm" list item "for parsing: 2." says that "//"
indicates the descendant axis. This is misleading, since you would
expect "//A/B" to match in the document "<A><B/></A>". However, if "//"
means "descendant", this results in
"/descendant::node()/child::A/child::B", and since there is no node
between <A> and the document root, this XPath would evaluate to {}.

This should close my ACTION-614.

best regards


Dipl.-Inf. Meiko Jensen
Chair for Network and Data Security 
Horst Görtz Institute for IT-Security 
Ruhr University Bochum, Germany
Universitätsstr. 150, Geb. IC 4/150
D-44780 Bochum, Germany
Phone: +49 (0) 234 / 32-26796
Telefax: +49 (0) 234 / 32-14347
http:// www.nds.rub.de
Received on Monday, 9 August 2010 15:10:30 UTC

This archive was generated by hypermail 2.3.1 : Tuesday, 6 January 2015 20:55:14 UTC