Denial Of Service mitigations#

There are multiple ways a client can overload a server and deny service to other clients. A simple example can simply be to call an expensive route at a rate high enough that the server’s resources are exhausted. Whilst this could happen innocently (if a route was popular) it takes a lot of client resource, there are malicious methods to exhaust server resources.

There are two attack methods mitigated and discussed here, the first aims to open as many connections to the server as possible without freeing them, thereby eventually exhausting all the connections and preventing other clients from connecting. As most request response cycles last milliseconds before the connection is closed the key is to somehow hold the connection open.

The second aims to exhaust the server’s memory and hence either slow the server to a crawl or kill the server application, thereby preventing the server from replying to other clients. The key here is to somehow make the server write a lot of information to memory.

Inactive connection#

This attack is of the first type and aims to exhaust the server’s connections. It works by opening connections to the server and then doing nothing with the connection. A poorly configured server would simply wait for the client to do something therefore holding the connection open.

To mitigate this Hypercorn has a keep alive timeout that will keep an inactive connection alive for this time. It can be configured via the configuration keep_alive_timeout setting.

The default value for the keep alive timeout is 5 seconds which is the max recommended default in the Gunicorn settings.

Note

Connections are not considered inactive whilst the request is being processed. So this will only timeout connections inactive before a request, or after a response and before the next request. This does not affect websocket connections.

Large request body#

This attack is of the second type and aims to exhaust the server’s memory by inviting it to receive a large request body (and hence write the body to memory). A poorly configured server would have no limit on the request body size and potentially allow a single request to exhaust the server.

It is up to the framework to guard against this attack. This is to allow the framework to consume the request body if desired.

Slow request body#

This attack is of the first type and aims to exhaust the server’s connections by inviting it to wait a long time for the request’s body. A poorly configured server would wait indefinitely for the request body.

It is up to the framework to guard against this attack. This is to allow the framework to consume the request body if desired.

No response consumption#

This attack is of the second type and aims to exhaust the server’s memory by failing to consume the data sent to the client. This failure results in backpressure on the server that leads to the response being written to memory rather than the connection. A poorly configured server would ignore the backpressure and exhaust its memory. (Note this requires a route that respondes with a lot of data, e.g. video streaming).

To mitigate this Hypercorn responds to the backpressure and pauses (blocks) the coroutine writing the response.

Slow response consumption#

This attack is of the first type and aims to exhaust the server’s connections by inviting the server to take a long time sending the response, for example by applying backpressure indefinetly. A poorly configured server would simply wait indefinetly trying to send the response.

It is up to the framework to guard against this attack. This is to allow for responses that purposely take a long time, e.g. server sent events.

Large websocket message#

This attack is of the second type and aims to exhaust the server’s memory by inviting it to receive very large websocket messages. A poorly configured server would have no limit on the message size and potentially allow a single message to exhaust the server.

To mitigate this Hypercorn limits the message size to the value set in the application config websocket_max_message_size. Any message larger than this limit will trigger the websocket to be closed abruptly.

The default value for websocket_max_message_size is 16 MB, which is chosen as it is the limit discussed in the Flask documentation.

Slow SSL handshake#

This attack is of the first type and aims to exhaust the server’s connections by failing to complete (or potentially even start) the SSL handshake. A poorly configured server would simply wait indefinetly for the SSL handshake to complete.

To mitigate this Hypercorn has a ssl handshake timeout that will close connections that take longer to complete the ssl handshake. It can be configured via the configuration ssl_handshake_timeout setting.

The default value for ssl_handshake_timeout is 60 seconds, which is chosen as it is the limit used in NGINX.

HTTP/2 attacks#

There are a number of specific attacks against the HTTP/2 protocol that are designed to exhuast the server’s resources. These are only feasible against a server serving over HTTP/2, and hence apply to Hypercorn by default.

Flood attacks#

This attack aims to exhaust the server’s resources by requesting ack frames whilst not consuming them. It is conceptually similar to the “No response consumption” attack discussed above however this attack slowly consumes frames that are unrelated to the HTTP response such as ping, setting, and reset frames.

To mitigate this Hypercorn responds to the backpressure and pauses (blocks) the reading data from the client. Depending on the state of the connection a timeout should then close the connection.

Data Dribble#

This attack aims to exhaust the server’s resources by requesting a response whilst setting the window size to a very small value e.g. 1 byte. The server could then exhaust memory buffering the response. It is conceptually similar to the “No response consumption” attack discussed above.

It is up to the framework to guard against this attack. This is to allow for responses that purposely take a long time, e.g. server sent events.

Internal Data Buffering#

This attack is conceptually similar to the “Flood attack” and “No response consumption” attack in that it requires the server to buffer data that it cannot send to the client.

To mitigate this Hypercorn responds to the backpressure and pauses (blocks) the coroutine writing the response.