First cleanup/remove of option.to_dict

[karlsson]

Add option.merge_type_list
Use Count(n) instead of Once in ActiveState
Add handling of tcp_error, ssl_error, tcp_passive and ssl_passive

[rawhat]

In testing this locally, the active mode changes are pretty significantly worse for performance. A simple "hello, world" server saw the performance drop from ~300k RPS to ~250k RPS. This was repeatable.

Restoring the {active, once} behavior actually improved performance, up to ~310k RPS. So I think the other changes are definitely worth including here.

[rawhat]

just some random thoughts

[karlsson]

I will revert to {active, once} but I cannot see any reason for setting it when you receive User(msg). My understanding is that it is only passivated when receiving on the socket, but maybe I am missing something?
I tried to only set it upon receiving Internal(Receive(msg)) and all tests pass, also for ewe.

Update: Ok, I get it, in case the user provides a user selector with handlers that overrides the "base" selector.

[karlsson]

In the handler:

case msg {
  ....
  Internal(Closed) | Internal(Close) ->

from where do Internal(Close) originate?
I can see Closed from ssl_closed and tcp_closed .

[vshakitskiy]

In the handler:

case msg {
  ....
  Internal(Closed) | Internal(Close) ->

from where do Internal(Close) originate? I can see Closed from ssl_closed and tcp_closed .

As far as I know, mist uses Internal(Close) to handle timer timeout.
https://github.com/rawhat/mist/blob/master/src/mist/internal/http/handler.gleam#L109

[karlsson]

Oh, Close message is part of the "external" API. Thanks.
Maybe it could be made more explicit in a public function or part of configuration.

[karlsson]

I have made it possible to set ActiveState for the TCP server to Active and Count(N) also (besides the old default Once). But I have problems to make all mist and ewe test cases pass with the new configurations. For mist some "100-continue" testcase fails since the handler loop returns NormalStop on the first chunk and I get NormalStop on ewe fragmented tests. All received messages looks the same for both Once and Active though so I am rather confused.

[vshakitskiy]

I will try to play around tomorrow and let you know if I figure something out on ewe side :)

[vshakitskiy]

Quoting my discord message:

So, the reason my fragmented tests are failing, is because I am using transport.receive_timeout to read from socket. And, as I figured out, gen_tcp:recv and ssl:recv cannot be used on active sockets, as messages sent through a connection are automatically read from the socket for us and placed in the controlling process's mailbox, so I assume receive functions return {error, einval}

[karlsson]

Thanks for the investigation! I will check the mist case, could be similar cause there since it concerns a "split" message. Nothing wrong really, just limited to ActiveState(Once) only.

Update: the mist case also involves reading data directly from the socket for split messages. Nothing wrong, just limited to ActiveState(Once) configuration.

I will set the PR "Ready for review".

[vshakitskiy]

Looks good to me, great job!