Included in this folder is a complete ThousandIsland.Handler
based implementation of HTTP/2 as
defined in RFC 7540.
Within a Bandit server, an HTTP/2 connection is modeled as a set of processes:
- 1 process per connection, a
Bandit.HTTP2.Handler
module implementing theThousandIsland.Handler
behaviour, and; - 1 process per stream (i.e.: per HTTP request) within the connection, implemented as
a
Bandit.HTTP2.StreamTask
Task
The lifetimes of these processes correspond to their role; a connection process lives for as long as a client is connected, and a stream process lives only as long as is required to process a single stream request within a connection.
Connection processes are the 'root' of each connection's process group, and are supervised by
Thousand Island in the same manner that ThousandIsland.Handler
processes are usually supervised
(see the project README for details).
Stream processes are not supervised by design. The connection process starts new stream processes as required, and does so
once a complete header block for a new stream has been received. It starts stream processes via
a standard start_link
call, and manages the termination of the resultant linked stream processes
by handling {:EXIT,...}
messages as described in the Elixir documentation. This approach is
aligned with the realities of the HTTP/2 model, insofar as if a connection process terminates
there is no reason to keep its constituent stream processes around, and if a stream process dies
the connection should be able to handle this without itself terminating. It also means that our
process model is very lightweight - there is no extra supervision overhead present because no such
supervision is required for the system to function in the desired way.
The overall structure of the implementation is managed by the Bandit.HTTP2.Handler
module, and
looks like the following:
- Bytes are asynchronously received from ThousandIsland via the
Bandit.HTTP2.Handler.handle_data/3
function - Frames are parsed from these bytes by calling the
Bandit.HTTP2.Frame.deserialize/2
function. If successful, the parsed frame(s) are returned. We retain any unparsed bytes in a buffer in order to attempt parsing them upon receipt of subsequent data from the client - Parsed frames are passed into the
Bandit.HTTP2.Connection
module along with a struct of same module. Frames are applied against this struct in a vaguely FSM-like manner, using pattern matching within theBandit.HTTP2.Connection.handle_frame/3
function. Any side-effects of received frames are applied in these functions, and an updated connection struct is returned to represent the updated connection state. These side-effects can take the form of starting stream tasks, conveying data to running stream tasks, responding to the client with various frames, or any number of other actions - This process is repeated every time we receive data from the client until the
Bandit.HTTP2.Connection
module indicates that the connection should be closed, either normally or due to error. Note that frame deserialization may end up returning a connection error if the parsed frames fail specific criteria (generally, the frame parsing modules are responsible for identifying errors as described in section 6 of RFC 7540). In these cases, the failure is passed through to the connection module for processing in order to coordinate an orderly shutdown or client notification as appropriate
The details of a particular stream are contained within a Bandit.HTTP2.Stream
struct
(as well as a Bandit.HTTP2.StreamTask
process in the case of active streams). The
Bandit.HTTP2.StreamCollection
module manages a collection of streams, allowing for the memory
efficient management of complete & yet unborn streams alongside active ones.
Once a complete header block has been read, a Bandit.HTTP2.StreamTask
is started to manage the
actual calling of the configured Plug
module for this server, using the Bandit.HTTP2.Adapter
module as the implementation of the Plug.Conn.Adapter
behaviour. This adapter uses a simple
receive
pattern to listen for messages sent to it from the connection process, a pattern chosen
because it allows for easy provision of the blocking-style API required by the Plug.Conn.Adapter
behaviour. Functions in the Bandit.HTTP2.Adapter
behaviour which write data to the client use
GenServer
calls to the Bandit.HTTP2.Handler
module in order to pass data to the connection
process.
All of this is exhaustively tested. Tests are broken up primarily into protocol_test.exs
, which
is concerned with aspects of the implementation relating to protocol conformance and
client-facing concerns, while plug_test.exs
is concerned with aspects of the implementation
having to do with the Plug API and application-facing concerns. There are also more
unit-style tests covering frame serialization and deserialization.
In addition, the h2spec
conformance suite is run via a System
wrapper & executes the entirety
of the suite (in strict mode) against a running Bandit server.
Some limitations and assumptions of this implementation:
- This handler assumes that the HTTP/2 connection preface has already been consumed from the
client. The
Bandit.InitialHandler
module uses this preface to discriminate between various HTTP versions when determining which handler to use - Priority frames are parsed and validated, but do not induce any action on the part of the server. There is no priority assigned to respective streams in terms of processing; all streams are run in parallel as soon as they arrive
- While flow control is completely implemented here, the specific values used for upload flow
control (that is, the end that we control) are fixed. Specifically, we attempt to maintain
fairly large windows in order to not restrict client uploads (we 'slow-start' window changes
upon receipt of first byte, mostly to retain parity between connection and stream window
management since connection windows cannot be changed via settings). The majority of flow
control logic has been encapsulated in the
Bandit.HTTP2.FlowControl
module should future refinement be required