Caddy also has this bug, but I'm assuming that's because they use net/http.

invalid because it is missing its Content-Length header:

Is a Content-Length header field actually required for a response that specifies Connection: close? (What leads you to believe that it is invalid?)

RFC 9110 says:

A user agent SHOULD send Content-Length in a request when the method defines a meaning for enclosed content and it is not sending Transfer-Encoding.

(Note that that is “SHOULD”, not “MUST”.)

And RFC 9112 says:

Otherwise [if no Content-Length or Transfer-Encoding header field is present], this is a response message without a declared message body length, so the message body length is determined by the number of octets received prior to the server closing the connection.

As far as I can tell, per RFC 9112 this is a valid HTTP/1.1 response.

(CC @neild)

I concur.

In addition, RFC 9110 says:

[...] in the absence of Transfer-Encoding, an origin server SHOULD send a Content-Length header field when the content size is known prior to sending the complete header section. This will allow downstream recipients to measure transfer progress, know when a received message is complete, and potentially reuse the connection for additional requests.

This response is small enough that measuring progress is unnecessary (the response should fit in a single packet), and the connection is closed immediately after the response so connection reuse is impossible. It would not be incorrect for us to send a Content-Length here, but I believe this is a valid HTTP/1.1 response.

Thank you for the quick responses.

I agree that the spec only suggests the inclusion of a TE or CL header; I was mistaken.

That said, net/http is the only HTTP library of which I am aware that does not send a TE or CL header in its responses.

Among those that do send TE or CL headers in 400 responses are

• AIOHTTP
• Apache HTTPD
• Boost::Beast
• Gunicorn
• H2O
• IIS
• Jetty
• Ligghtpd
• NGINX
• Node.js
• OpenLitespeed
• Puma
• Tomcat
• Tornado

Given that the spec says that you SHOULD do this, I stand by my suggestion that we move away from net/http's current behavior.

EDIT: Typo

What's the benefit in sending Content-Length here?

If there are existing broken client implementations in common use that can't cope with this response, then that might be an argument for change, but I'm dubious that this particular form of buggy client is common.

The benefit of sending CL is for clarity and ease of parsing.

This is the top answer to a StackOverflow question asking what happens when an HTTP response has no CL header:

In HTTP/1.0 - server response without content-length is when the stream closes
In HTTP/1.1 - server response without content-length is when the response is chunked encoded

Believing this answer would lead to exactly the sort of bug you describe.

That said, I think I'm convinced that the best course of action is not to add a CL header, but to remove the message bodies from the default 400 responses. They're redundant because of the reason string, and removing them should improve performance marginally, and improve compatibility with other servers.

The Stack Overflow response is wrong.

In the absence of evidence that the current behavior is causing real world problems, I believe that not only is there no good reason to change it, but there is a small but real value in preserving it as a minor discouragement to the development of new, buggy HTTP/1.1 client implementations that cannot handle this valid response.

I am aware that the SO response is wrong; I sent it to demonstrate that my misunderstanding of the standard is a widespread one.

If you'd like net/http to be the lone upholder of a little-known part of the HTTP standard, then that is your choice, and I will close the issue.

Here's another portion of the standard that encourages being explicit about lengths, from RFC 9112 Section 6.3:

Since there is no way to distinguish a successfully completed, close-delimited response message from a partially received message interrupted by network failure, a server SHOULD generate encoding or length-delimited messages whenever possible. The close-delimiting feature exists primarily for backwards compatibility with HTTP/1.0.

Thus, a client will have no way of determining with reasonable certainty that it received the entirety of net/http's error response when it receives a response with no Content-Length.