Parser generates invalid URLs

[LEW21]

https://url.spec.whatwg.org/commit-snapshots/a1b789c6b6c36fcdb16311da5abd177e84151fca/#url-parsing

For each byte in buffer:​

If byte is less than 0x21 (!), greater than 0x7E (~), or is 0x22 ("), 0x23 (#), 0x3C (<), or 0x3E (>), append byte, percent encoded, to url’s query.

Otherwise, append a code point whose value is byte to url’s query.

This leads to creation of invalid URLs - ones that contain [, \, ], ^, `, {, |, }, which are neither URL code points nor '%' and trigger validation errors:

Otherwise:

If c is not a URL code point and not U+0025 (%), validation error.

I think that either:

• the list of valid query characters should be expanded to include more characters, or
• these ones should be escaped too.

Related issues: #378, #17

[LEW21]

Actually, there is a similar problem for other URL parts, too.

For example:

• generated path may contain [, \, ], ^, | - which are not valid there;
• generated fragment may contain the above and #, {, } - which are not valid there;
• if state override is given, valid path may contain ? (which will be then escaped)

[LEW21]

I'd also consider excluding ` from the fragment percent-encode set, like it was excluded from query in #17, so that fragment percent-encode set is contained within query percent-encode set.

[annevk]

I don't think we're in a position to change these, unless implementations vary enough for a particular code point that there's some leeway.

What we should maybe do is document the when the parser doesn't "fixup" an invalid URL more clearly.

[domenic]

I think it would definitely be worth attempting to change these if at all possible, or else expanding the definition of valid URL.

[annevk]

I don't think we can change most of these due to compatibility, but if someone wants to have another attempt at doing the research I'm willing to assist.

I don't think we should change the definition of what's valid either though. Validity in part helps signaling problems you might face, including with legacy software. (Same as with HTML.)

[domenic]

It seems extraordinarily strange that we're not only giving developers the tools to create invalid URLs, but we're also encouraging other standards to produce invalid URLs and pass them on to the rest of the ecosystem (e.g., over HTTP). In that case I'd question why we're calling such URLs invalid at all. At that point they're just "URLs produced by all software that follows the URL Standard", and validity doesn't buy us much.

I don't believe you can parse/serialize any string through the HTML parser and get invalid HTML, for example. (Maybe some edge cases exist, but these characters are hardly edge cases.)

[annevk]

Other standards? I don't see how this is different from #118.

[domenic]

Sure. Every standard that uses the URL parser on user input is currently producing invalid URLs, right? Including standards that then use the URL serializer and send the result to the network or elsewhere.

#118 stemmed from suggestions that the parser should fail when given an invalid URL. This issue is a concern about the parser producing an invalid URL.

[annevk]

@domenic they can produce invalid URLs, sure. Just like the HTML parser can produce invalid HTML.

[annevk]

#118 was precisely about this. That the syntax (valid URLs) conflicts with the parser.

[LEW21]

@annevk, I'd agree that this is a case of Garbage-In-Garbage-Out, if it would only happen when parsing invalid URLs. However, it also happens when changing components of an URL object:

x = new URL('http://localhost')
x.search = 'a#{}'
x.href // "http://localhost/?a%23{}"

Also, the browsers don't 100% follow the spec here - for example Chrome escapes ^ and | in paths, while the standard says it should not. I'll do more tests to check how browsers behave on all chars in all places.

[LEW21]

OK, these are the results for Chrome, Firefox, Edge: https://docs.google.com/spreadsheets/d/1mSl2N2Wrc7ZdKy2ArhLg0t3EHI2DOdHaDyWMZQByHE4/edit?usp=sharing

Note: I'm not sure if I've set spec behavior of \ correctly.

I was testing with:

U = () => new URL('http://localhost')
s = 'x "#%<>?[\\]^`{|}$'
u = U(); u.pathname = s; u.href
u = U(); u.search = s; u.href
u = U(); u.hash = s; u.href

On Edge, u.pathname = s and u.search = s were throwing errors, so I had to check these character-by-character with loops:

for (let c of s) {try { u = U(); u.pathname = 'x' + c + 'x'; console.log(c, u.href); } catch (e) { console.log(c, e) }}
for (let c of s) {try { u = U(); u.search   = 'x' + c + 'x'; console.log(c, u.href); } catch (e) { console.log(c, e) }}

[LEW21]

Looking at the results, I think that:

• ^ should be added to the path percent-encode set
• ` should be moved from the fragment percent-encode set to the path percent-encode set
• path, query and fragment should treat [, ] and | as valid
• query and fragment should treat ?, \, ^, `, { and } as valid
• fragment should treat # as valid

The only confusing one is:

• \ in paths - no idea what to do here, I don't really understand the backslash-magic in the spec