`websocket: [fullsweep_after: ...]` not applied to websocket processes with high memory use

[balysv]

Environment

erlang: 24.3.4
elixir: 1.13.0
Cowboy: 2.9.0
plug_cowboy: 2.5.2
phoenix: 1.6.9
OS: Ubuntu 22.04

Related issue: https://github.com/phoenixframework/phoenix/issues/4613

Expected behavior

Setting websocket: [fullsweep_after: 0] on sockets should ideally apply the process flag to the {:cowboy_websocket, :loop, 3} process that has high memory usage and often needs a forced garbage collection.

Actual behavior

When dealing with very large payloads {:cowboy_websocket, :loop, 3} processes accumulate memory usage that lead to OOM errors.

fullsweep_after flag introduced here is only applied to children of the offending processes. Running manual GCs of said processes temporarily fixes the issue.

Inspecting process tree at the time :fullsweep_after is applied:

[
  ...
  dictionary: [
     "$initial_call": {:cowboy_stream_h, :request_process, 3},
     "$ancestors": [#PID<0.5297.0>, ...]
  ],
  garbage_collection: [
    fullsweep_after: 0,
    ...
  ]
]

Parent (the one with high memory usage)

iex(14)> Process.info(pid(0, 5297, 0))
[
  current_function: {:cowboy_websocket, :loop, 3},
  dictionary: [
    "$initial_call": {:cowboy_clear, :connection_process, 4},
    "$ancestors": [...]
  ],
  trap_exit: true,
  garbage_collection: [
    fullsweep_after: 65535, <<< would like to have this overriden to 0
    ...
  ]
]

Any other suggestions on how to deal with this without manual intervention are welcome.