HTTP/2 guide

The HTTP/2 protocol

HTTP/2 is the evolution of the world's most successful application layer protocol, HTTP. It focuses on making more efficient use of network resources. It does not change the fundamentals of HTTP, the semantics. There are still request and responses and headers and all that. So, if you already know HTTP/1, you know 95% about HTTP/2 as well.

There has been a lot written about HTTP/2 and how it works. The most normative is, of course, its RFC 7540 (also available in more readable formatting, YMMV). So, there you'll find the nuts and bolts.

But, as RFC do, it's not really a good thing to read first. It's better to first understand what a thing wants to do and then read the RFC about how it is done. A much better document to start with is http2 explained by Daniel Stenberg, the author of curl. It is available in an ever growing list of languages, too!

Too Long, Didn't read: there are some new terms and gotchas that need to be kept in mind while reading this document:

• HTTP/2 is a binary protocol, as opposed to HTTP 1.1 that is plain text. The latter is meant to be human readable (for example sniffing network traffic) meanwhile the former is not. More info in the official FAQ question.
• h2 is HTTP/2 over TLS (protocol negotiation via ALPN).
• h2c is HTTP/2 over TCP.
• A frame is the smallest unit of communication within an HTTP/2 connection, consisting of a header and a variable-length sequence of octets structured according to the frame type. More info in the official documentation section.
• A stream is a bidirectional flow of frames within the HTTP/2 connection. The correspondent concept in HTTP 1.1 is a request/response message exchange. More info in the official documentation section.
• HTTP/2 is able to run multiple streams of data over the same TCP connection, avoiding the classic HTTP 1.1 head of blocking slow request and avoiding to re-instantiate TCP connections for each request/response (KeepAlive patched the problem in HTTP 1.1 but did not fully solve it).

HTTP/2 in Apache httpd

The HTTP/2 protocol is implemented by its own httpd module, aptly named mod_http2. It implements the complete set of features described by RFC 7540 and supports HTTP/2 over cleartext (http:), as well as secure (https:) connections. The cleartext variant is named 'h2c', the secure one 'h2'. For h2c it allows the direct mode and the Upgrade: via an initial HTTP/1 request.

One feature of HTTP/2 that offers new capabilities for web developers is Server Push. See that section on how your web application can make use of it.

Build httpd with HTTP/2 support

mod_http2 uses the library of nghttp2 as its implementation base. In order to build mod_http2 you need at least version 1.2.1 of libnghttp2 installed on your system.

When you ./configure your Apache httpd source tree, you need to give it '--enable-http2' as additional argument to trigger the build of the module. Should your libnghttp2 reside in an unusual place (whatever that is on your operating system), you may announce its location with '--with-nghttp2=<path>' to configure.

While that should do the trick for most, they are people who might prefer a statically linked nghttp2 in this module. For those, the option --enable-nghttp2-staticlib-deps exists. It works quite similar to how one statically links openssl to mod_ssl.

Speaking of SSL, you need to be aware that most browsers will speak HTTP/2 only on https: URLs, so you need a server with SSL support. But not only that, you will need a SSL library that supports the ALPN extension. If OpenSSL is the library you use, you need at least version 1.0.2.

Basic Configuration

When you have a httpd built with mod_http2 you need some basic configuration for it becoming active. The first thing, as with every Apache module, is that you need to load it:

LoadModule http2_module modules/mod_http2.so

The second directive you need to add to your server configuration is

Protocols h2 http/1.1

This allows h2, the secure variant, to be the preferred protocol on your server connections. When you want to enable all HTTP/2 variants, you simply write:

Protocols h2 h2c http/1.1

Depending on where you put this directive, it affects all connections or just the ones to a certain virtual host. You can nest it, as in:

Protocols http/1.1
<VirtualHost ...>
    ServerName test.example.org
    Protocols h2 http/1.1
</VirtualHost>

This allows only HTTP/1 on connections, except SSL connections to test.example.org which offer HTTP/2.

The order of protocols mentioned is also relevant. By default, the first one is the most preferred protocol. When a client offers multiple choices, the one most to the left is selected. In

Protocols http/1.1 h2

the most preferred protocol is HTTP/1 and it will always be selected unless a client only supports h2. Since we want to talk HTTP/2 to clients that support it, the better order is

Protocols h2 h2c http/1.1

There is one more thing to ordering: the client has its own preferences, too. If you want, you can configure your server to select the protocol most preferred by the client:

ProtocolsHonorOrder Off

makes the order you wrote the Protocols irrelevant and only the client's ordering will decide.

A last thing: the protocols you configure are not checked for correctness or spelling. You can mention protocols that do not exist, so there is no need to guard Protocols with any <IfModule> checks.

For more advanced tips on configuration, see the modules section about dimensioning and how to manage multiple hosts with the same certificate.

MPM Configuration

HTTP/2 is supported in all multi-processing modules that come with httpd. However, if you use the prefork mpm, there will be severe restrictions.

In prefork, mod_http2 will only process one request at at time per connection. But clients, such as browsers, will send many requests at the same time. If one of these takes long to process (or is a long polling one), the other requests will stall.

mod_http2 will not work around this limit by default. The reason is that prefork is today only chosen, if you run processing engines that are not prepared for multi-threading, e.g. will crash with more than one request.

If your setup can handle it, configuring event mpm is nowadays the best one (if supported on your platform).

If you are really stuck with prefork and want multiple requests, you can tweak the H2MinWorkers to make that possible. If it breaks, however, you own both parts.

Clients

Almost all modern browsers support HTTP/2, but only over SSL connections: Firefox (v43), Chrome (v45), Safari (since v9), iOS Safari (v9), Opera (v35), Chrome for Android (v49) and Internet Explorer (v11 on Windows10) (source).

Other clients, as well as servers, are listed on the Implementations wiki, among them implementations for c, c++, common lisp, dart, erlang, haskell, java, nodejs, php, python, perl, ruby, rust, scala and swift.

Several of the non-browser client implementations support HTTP/2 over cleartext, h2c. The most versatile being curl.

Useful tools to debug HTTP/2

The first tool to mention is of course curl. Please make sure that your version supports HTTP/2 checking its Features:

    $ curl -V
    curl 7.45.0 (x86_64-apple-darwin15.0.0) libcurl/7.45.0 OpenSSL/1.0.2d zlib/1.2.8 nghttp2/1.3.4
    Protocols: dict file ftp ftps gopher http https imap imaps ldap ldaps pop3 [...] 
    Features: IPv6 Largefile NTLM NTLM_WB SSL libz TLS-SRP HTTP2
    

And for really deep inspection wireshark.

The nghttp2 package also includes clients, such as:

• nghttp - useful to visualize the HTTP/2 frames and get a better idea of the protocol.
• h2load - useful to stress-test your server.

Chrome offers detailed HTTP/2 logs on its connections via the special net-internals page. There is also an interesting extension for Chrome and Firefox to visualize when your browser is using HTTP/2.

Server Push

The HTTP/2 protocol allows the server to PUSH responses to a client it never asked for. The tone of the conversation is: "here is a request that you never sent and the response to it will arrive soon..."

But there are restrictions: the client can disable this feature and the server may only ever PUSH on a request that came from the client.

The intention is to allow the server to send resources to the client that it will most likely need: a css or javascript resource that belongs to a html page the client requested. A set of images that is referenced by a css, etc.

The advantage for the client is that it saves the time to send the request which may range from a few milliseconds to half a second, depending on where on the globe both are located. The disadvantage is that the client may get sent things it already has in its cache. Sure, HTTP/2 allows for the early cancellation of such requests, but still there are resources wasted.

To summarize: there is no one good strategy on how to make best use of this feature of HTTP/2 and everyone is still experimenting. So, how do you experiment with it in Apache httpd?

mod_http2 inspect response header for Link headers in a certain format:

Link </xxx.css>;rel=preload, </xxx.js>; rel=preload

If the connection supports PUSH, these two resources will be sent to the client. As a web developer, you may set these headers either directly in your application response or you configure the server via

<Location /xxx.html>
    Header add Link "</xxx.css>;rel=preload"
    Header add Link "</xxx.js>;rel=preload"
</Location>

If you want to use preload links without triggering a PUSH, you can use the nopush parameter, as in

Link </xxx.css>;rel=preload;nopush

or you may disable PUSHes for your server entirely with the directive

H2Push Off

And there is more:

The module will keep a diary of what has been PUSHed for each connection (hashes of URLs, basically) and will not PUSH the same resource twice. When the connection closes, this information is discarded.

There are people thinking about how a client can tell a server what it already has, so PUSHes for those things can be avoided, but this is all highly experimental right now.

Another experimental draft that has been implemented in mod_http2 is the Accept-Push-Policy Header Field where a client can, for each request, define what kind of PUSHes it accepts.

PUSH might not always trigger the request/response/performance that one expects or hopes for. There are various studies on this topic to be found on the web that explain benefits and weaknesses and how different features of client and network influence the outcome. For example: just because the server PUSHes a resource does not mean a browser will actually use the data.

The major thing that influences the response being PUSHed is the request that was simulated. The request URL for a PUSH is given by the application, but where do the request headers come from? For example, will the PUSH request a accept-language header and if yes with what value?

Apache will look at the original request (the one that triggered the PUSH) and copy the following headers over to PUSH requests: user-agent, accept, accept-encoding, accept-language, cache-control.

All other headers are ignored. Cookies will also not be copied over. PUSHing resources that require a cookie to be present will not work. This can be a matter of debate. But unless this is more clearly discussed with browser, let's err on the side of caution and not expose cookie where they might ordinarily not be visible.

Early Hints

An alternative to PUSHing resources is to send Link headers to the client before the response is even ready. This uses the HTTP feature called "Early Hints" and is described in RFC 8297.

In order to use this, you need to explicitly enable it on the server via

H2EarlyHints on

(It is not enabled by default since some older browser tripped on such responses.)

If this feature is on, you can use the directive H2PushResource to trigger early hints and resource PUSHes:

<Location /xxx.html>
    H2PushResource /xxx.css
    H2PushResource /xxx.js
</Location>

This will send out a "103 Early Hints" response to a client as soon as the server starts processing the request. This may be much early than the time the first response headers have been determined, depending on your web application.

If H2Push is enabled, this will also start the PUSH right after the 103 response. If H2Push is disabled however, the 103 response will be send nevertheless to the client.