Title: A better algorithm for out-of-sockets eviction
Author: Nick Mathewson
Created: 25 July 2022
Status: Open


Our existing algorithm for handling an out-of-sockets condition needs improvement. It only handles sockets used for OR connections, and prioritizes those with more circuits. Because of these weaknesses, the algorithm is trivial to circumvent, and it's disabled by default with DisableOOSCheck.

Here we propose a new algorithm for choosing which connections to close when we're out of sockets. In summary, the new algorithm works by deciding which kinds of connections we have "too many" of, and then by closing excess connections of each kind. The algorithm for selecting connections of each kind is different.

Intuitions behind the algorithm below

We want to keep a healthy mix of connections running; favoring one kind of connection over another gives the attacker a fine way to starve the disfavored connections by making a bunch of the favored kind.

The correct mix of connections depends on the type of service we are providing. Everywhere except authorities, for example, inbound directory connections are perfectly fine to close, since nothing in our protocol actually generates them.

In general, we would prefer to close DirPort connections, then Exit connections, then OR connections.

The priority with which to close connections is different depending on the connection type. "Age of connection" or "number of circuits" may be a fine metric for how truly used an OR connection is, but for a DirPort connection, high age is suspicious.

The algorithm

Define a "candidate" connection as one that has a socket, and is either an exit stream, an inbound directory stream, or an OR connection.

(Note that OR connections can be from clients, relays, or bridges. Note that ordinary relays should not get directory streams that use sockets, since clients always use BEGIN_DIR to create tunneled directory streams.)

In all of the following, treat subtraction as saturating at zero. In other words, when you see "A - B" below, read it as "MAX(A-B, 0)".

Phase 1: Deciding how many connections to close

When we find that we are low on sockets, we pick a number of sockets that we want to close according to our existing algorithm. (That is, we try to close 1/4 of our maximum sockets if we have reached our upper limit, or 1/10 of our maximum sockets if we have encountered a failure from socket(2).) Call this N_CLOSE.

Then we decide which sockets to target based on this algorithm.

  1. Consider the total number of sockets used for exit streams (N_EXIT), the total number used for inbound directory streams (N_DIR), and the total number used for OR connections (N_OR). (In these calculations, we exclude connections that are already marked to be closed.) Call the total N_CONN = N_DIR + N_OR + N_EXIT. Define N_RETAIN = N_CONN - N_CLOSE.

  2. Compute how many connections of each type are "in excess". First, calculate our target proportions:

    • If we are an authority, let T_DIR = 1. Otherwise set T_DIR = 0.1.
    • If we are an exit or we are running an onion service, let T_EXIT = 2. Otherwise let T_EXIT = 0.1.
    • Let T_OR = 1.

    TODO: Should those numbers be consensus parameters?

    These numbers define the relative proportions of connections that we would be willing to retain retain in our final mix. Compute a number of excess connections of each type by calculating.

    T_TOTAL = T_OR + T_DIR + T_EXIT.
    EXCESS_OR    = N_OR   - N_RETAIN * (T_OR   / T_TOTAL)
  3. Finally, divide N_CLOSE among the different types of excess connections, assigning first to excess directory connections, then excess exit connections, and finally to excess OR connections.


We will try to close CLOSE_DIR directory connections, CLOSE_EXIT exit connections, and CLOSE_OR OR connections.

Phase 2: Closing directory connections

We want to close a certain number of directory connections. To select our targets, we sort first by the number of directory connections from a similar address (see "similar address" below) and then by their age, preferring to close the oldest ones first.

This approach defeats "many requests from the same address" and "Open a connection and hold it open, and do so from many addresses". It doesn't do such a great job with defeating "open and close frequently and do so on many addresses."

Note that fallback directories do not typically use sockets for handling directory connections: theirs are usually created with BEGIN_DIR.

Phase 3: Closing exit connections.

We want to close a certain number of exit connections. To do this, we pick an exit connection at random, then close its circuit along with all the other exit connections on the same circuit. Then we repeat until we have closed at least our target number of exit connections.

This approach probabilistically favors closing circuits with a large number of sockets open, regardless of how long those sockets have been open. This defeats the easiest way of opening a large number of exit streams ("open them all on one circuit") without making the counter-approach ("open each exit stream on its own circuit") much more attractive.

Phase 3: Closing OR connections.

We want to close a certain number of OR connections, to clients, bridges, or relays.

To do this, we first close OR connections with zero circuits. Then we close all OR connections but the most recent 2 from each "similar address". Then we close OR connections at random from among those not to a recognized relay in the latest directory. Finally, we close OR connections at random.

We used to unconditionally prefer to close connections with fewer circuits. That's trivial for an adversary to circumvent, though: they can just open a bunch of circuits on their bogus OR connections, and force us to preferentially close circuits from real clients, bridges, and relays.

Note that some connections that seem like client connections ("not from relays in the latest directory") are actually those created by bridges.

What is "A similar address"?

We define two connections as having a similar address if they are in the same IPv4 /30, or if they are in the same IPv6 /90.


This proposal was inspired by a set of OOS improvements from starlight.