LIBXML_NSCLEAN and DTD entities with namespaced attributes

[alfsb]

While doing some optimization experiments on doc-base/configure.php, I tried to add LIBXML_NSCLEAN before first DOMDocument->load(). This caused 173 validating failures, all related to the href attribute on Jing. libxml2 simply refused the file. All failures are like this:

doc-base/temp/manual.xml:18673:62: error: attribute "href" not allowed here; expected attribute [...] or an attribute from another namespace

manual.xml:18673:62 is something like

 builds from <link href="https://www.php.net/downloads.php">the official download

This, in turn is part of DTD entity

<!ENTITY warn.install.third-party-support '<warning xmlns="http://docbook.org/ns/docbook" xmlns:xlink="http://www.w3.org/1999/xlink">
<simpara>
Builds from third-parties are considered unofficial and not directly
supported by the PHP project. Any bugs encountered should be reported to the
provider of those unofficial builds unless they can be reproduced using the
builds from <link xlink:href="&url.php.downloads;">the official download
area</link>.
</simpara>
</warning>'>

The only thing that caught my eye is that xmlns:xlink="http://www.w3.org/1999/xlink" is declared at the "root" of entity, whereas other entities declares this namespace in the <link> element itself. But changing the positions, or adding the namespace in both places changed nothing.

The first difference between XMLs assembled with and without LIBXML_NSCLEAN, are:

< builds from <link xlink:href="https://www.php.net/downloads.php">the official download
> builds from <link href="https://www.php.net/downloads.php">the official download

This means that the LIBXML_NSCLEAN option removed the XML declarations from <simpara> (correctly), but also removed the xmlns prefix of href attribute, that appears to be incorrect, IMHO.

I do not know the XML spec by the letter, so I do not know if it is expected or if I hit a bug.

Specifically, I do not know if the parsing of namespaces of attributes can or should be treated differently, when inside a DTD entity.

This is on PHP 8.1.2-1ubuntu2.24, libxml 2.9.13.

[alfsb]

To be exact, this won't affect the manual building as is, as it does not use LIBXML_NSCLEAN. And AFAIK, this also won't affect XML entities, as it avoids all code paths related to DTD entities.

The size difference in English XML, with and without LIBXML_NSCLEAN, is 44670436 vs 40893789, or about 3.6 Mb of redundant namespaces declarations, almost 8,4% off all bytes.

This is the expected savings in size for XML namespaces declarations alone, if XML entities replace all DTD entities one day.

[nwellnhof]

This is on PHP 8.1.2-1ubuntu2.24, libxml 2.9.13.

This version of libxml2 is rather old. As there have been some improvements lately, it would be interesting to see whether recent versions behave in a better way.

That said, namespaces in entities are still buggy in libxml2, see libxml2 issue #601. Unfortunately, i didn't find the time to fix this issue once and for all while I was maintaining this library.