why unix sockets

gistfile1.txt

//! This module provides implementations of `ActiveConnector` and `PassiveConnector` that are backed
//! by UNIX socket instead of the TCP sockets that are normally used.
//!
//! Using UNIX sockets instead of TCP ones serves multiple purpose:
//!
//! - creating and connecting to UNIX socket doesn't require specific permissions so using them in
//!   tests mean that non-root users can run the tests. With TCP, listening on some ports may
//!   require root privileges.
//! - We can easily create namespaces for the sockets used in each test by simply using the
//!   filesystem hierarchy which means we avoid the classic "address already in use" pitfall. For
//!   instance, we cannot have two tests that create a TCP listener on 0.0.0.0:0:179.
//!
//! Swapping TCP sockets with UNIX sockets required a few tricks the
//! testers might need to understand to debug their test.
//!
//! # Address mapping
//!
//! The first trick is address mapping. UNIX sockets use paths are an address, while TCP uses an
//! (ip, port) tuple. In order to hide the UNIX socket addresses from flock and from the tests, we
//! map from one addressing scheme to another using the following rules:
//!
//! - for a given test, the UNIX sockets are stored under
//!   `<UDS_SOCKET_PATH>/<process ID>/<test ID>/<VRF_kernel_intf_name>` where:
//!     - `<UDS_SOCKET_PATH>` is a constant
//!     - `<process ID` is the current process ID
//!     - `<test ID>` is an ID generated by each test and stored as a thread-local storage variable
//!     - '<VRF_kernel_intf_name>' is the kernel's VRF interface name for non-default VRF, for default
//!        VRF there is no intf name appended to the path
//! - a TCP address w.x.y.z:123 in vrf blue is mapped to
//!   `<UDS_SOCKET_PATH>/<process ID>/<test ID>/blue_kernel_intf_name/z.x.y.z:123`
//!   and vice-versa
//!
//! # Unspecified addresses
//!
//! ## Unspecified IP
//!
//! `0.0.0.0:179` is a valid TCP address to listen on, and it results in the listener listening for
//! connections on all the local addresses on port 179 in default vrf. We cannot really do this with
//! UNIX sockets, so such an address is simply mapped to
//! `<UDS_SOCKET_PATH>/<process ID>/<test ID>/0.0.0.0:179`,
//! and when connectors must connect to `0.0.0.0:179` directly.
//!
//! ## Unspecified port
//!
//! Sometimes, the port is left out. Typically, when initiating a connection, we usually let the OS
//! pick a port random number for the socket. We sort of emulated a very limited version of this in
//! this module. In each `<UDS_SOCKET_PATH>/<process ID>/<test ID>/VRF_kernel_intf_name` directory a
//! file called `last-generated-port-number` keeps track of the last port number that was generated.
//! Each a new socket is created and the port number is unspecified, the number in that file is
//! incremented and used. This is broken two majors ways:
//!
//! - the filesystem is notoriously racy, multiple threads could try to read/write this file
//! - there no guarantee the port number that is picked is not actually already used
//!
//! However, this seems to be good enough for our tests, so far ¯\_(ツ)_/¯