xmlns-prefixed attribute on root element is lost

[jdm]

<!-- before -->
<svg xmlns="http://www.w3.org/2000/svg" xmlns:sodipodi="http://inkscape.sourceforge.net/DTD/sodipodi-0.dtd">
    <sodipodi:namedview>
        ...
    </sodipodi:namedview>

    <path d="..." sodipodi:nodetypes="cccc"/>
</svg>

<!-- after: xmlns:sodipodi moved to child -->
<svg xmlns="http://www.w3.org/2000/svg">
    <sodipodi:namedview xmlns:sodipodi="http://inkscape.sourceforge.net/DTD/sodipodi-0.dtd">
        ...
    </sodipodi:namedview>

    <path d="..." sodipodi:nodetypes="cccc"></path>
</svg>

<!-- before -->
<svg xmlns="http://www.w3.org/2000/svg" xmlns:sodipodi="http://inkscape.sourceforge.net/DTD/sodipodi-0.dtd">
    <path d="..." sodipodi:nodetypes="cccc"/>
</svg>

<!-- after: xmlns:sodipodi removed entirely -->
<svg xmlns="http://www.w3.org/2000/svg">
    <path d="..." sodipodi:nodetypes="cccc"></path>
</svg>

Originally posted by @noahbald in #538 (comment)

[jdm]

It looks like the parsing code stores the namespace data for the node in the treebuilder's namespace stack, but the data about which tag declares a namespace is thrown away. Instead we store the namespace uri for child tags; this is why the xmlns-prefixed attribute jumps to the first child that uses the namespace.

[jdm]

Parsing code that strips out the xmlns attribute:

html5ever/xml5ever/src/tree_builder/mod.rs

Line 347 in feced06

self.declare_ns(attr);

Parsing code that stores the namespace URI on child tags:

html5ever/xml5ever/src/tree_builder/mod.rs

Line 297 in feced06

name.ns = ns_uri;

Serialization code that uses the namespace URI:

html5ever/xml5ever/src/serialize/mod.rs

Line 111 in feced06

self.find_or_insert_ns(name);

[jdm]

Testcase in rcdom/tests/xml-driver.rs:

#[test]
fn weird() {
    assert_serialization(
       r##"<svg xmlns="http://www.w3.org/2000/svg" xmlns:sodipodi="http://inkscape.sourceforge.net/DTD/sodipodi-0.dtd">
    <sodipodi:namedview>
        ...
    </sodipodi:namedview>

    <path d="..." sodipodi:nodetypes="cccc"/>
</svg>"##,
        driver::parse_document(RcDom::default(), Default::default())
            .from_utf8()
            .one(r##"<svg xmlns="http://www.w3.org/2000/svg" xmlns:sodipodi="http://inkscape.sourceforge.net/DTD/sodipodi-0.dtd">
    <sodipodi:namedview>
        ...
    </sodipodi:namedview>

    <path d="..." sodipodi:nodetypes="cccc"/>
</svg>"##.as_bytes()),
    );
}

[jdm]

Current status: I have changes that allow these testcases to pass. However, they cause a lot of xml5lib namespace tests to fail, and Ygg01/xml5lib-tests@3fb8d2b suggests that the current behaviour is intentional. I'm hoping to get in touch with @Ygg01 to better understand what the right behaviour here is.

[jdm]

Branch: https://github.com/servo/html5ever/compare/xmlns-attrs?expand=1

[Ygg01]

Current status: I have changes that allow these testcases to pass. However, they cause a lot of xml5lib namespace tests to fail, and Ygg01/xml5lib-tests@3fb8d2b suggests that the current behaviour is intentional. I'm hoping to get in touch with @Ygg01 to better understand what the right behaviour here is.

Hi Josh. To be honest I don't quite remember the motivation. It's been like 8 years.

The important question is does it goes against XML namespace spec? If not, just change the tests. I'm open to PRs on xml5lib-tests

[Ygg01]

@jdm what are the next steps to close this?

• Do you want me to fix the xml5lib-tests?
• Move xml5lib-tests to the Servo repository, so you can take control over it?
• Make PRs to Ygg01/xm5lib-tests for me to review?

[noahbald]

Hey if it helps I've found a few documents which will become invalid/broken after parsing and serializing and then viewing in Firefox. Perhaps these can help?

Arch Linux Logo
Blobs
Isometric Madness
tldr-pages Banner
Wikipedia Logo

[Ygg01]

@jdm I've updated the xml5lib tests. New commit. Might require submodule update.

Looking back I think I'm starting to remember why I did it like that. And why I was wrong.

1. When converting XML ⇾ tree, the prefixes aren't kept. In fact, they are completely arbitrary. They are like temporary variables that are discarded after parsing, e.g. svg:width = "10" becomes {http://www.w3.org/2000/svg}width = "10" in tree.
2. When converting namespaces, the namespace definitions could be used by simplistic libs, to match on prefix, by scanning for namespace definitions, which is something that should never be done. You search by namespaces, not prefix. I omitted them to prevent such use cases. Trying to be correct by design.
3. I didn't consider someone would serialize and deserialize and expect similar/same results.