Initial Eio port #256
Initial Eio port #256
Conversation
talex5
force-pushed
the
eio
branch
5 times, most recently
from
25d6e31
to
9c74a31
Compare
This switches capnp-rpc from Lwt to Eio. One particularly nice side effect of this is that `Service.return_lwt` has gone, as there is no distinction now between concurrent and non-concurrent service methods. Notes: - In this commit, everything is still using the "lwt" names to make the diff useful. In a future commit, this should be renamed. Also, some of the "unix" functions can be moved into the core library with Eio. This would likely be a good time to rename `capnp_rpc` to `capnp_rpc_protocol` or something, leaving the short name free for the main library. - Mirage support is gone. Ideally, the regular library should work with Mirage 5.
|
Can we keep the LWT version available for OCaml 4.14? We will need to move services to OCaml 5 and EIO over a longer period of time, both for platform support reasons and just because it will take some time to port things over. |
The existing release will continue to work there. capnp-rpc is pretty stable, so I'm not expecting any major new features that would need back-porting. |
|
We'll take a release of this when it's ready. clarke and solver-service could use it. |
It executes code while including it!
Should probably buffer up messages instead, but this solves the immediate problem.
unix/network.ml
Outdated
| Eio_unix.Resource.fd_opt flow | ||
| |> Option.iter (fun fd -> | ||
| Eio_unix.Fd.use_exn "TCP_NODELAY" fd @@ fun fd -> | ||
| Unix.setsockopt fd Unix.TCP_NODELAY true |
There was a problem hiding this comment.
We might need to wrap this to catch
Unix.Unix_error(Unix.EOPNOTSUPP, "setsockopt", "")Stumbled across this new change when using this branch on macOS with a unix domain socket.
There was a problem hiding this comment.
Sounds good. Not sure whether to do it here, or get ocaml-multicore/eio#575 finished. We could then do something like:
Eio.Net.connect ~sw net eio_addr ~options:[Try (Delay false)]
capnp-rpc-net/endpoint.ml
Outdated
| let pp_error f = function | ||
| | `Closed -> Fmt.string f "Connection closed" | ||
| | `Msg m -> Fmt.string f m | ||
| Eio.Flow.shutdown t.flow `All |
There was a problem hiding this comment.
Wrapping this in a try stops us crashing when a flow has already been closed. I.e.
try
Eio.Flow.shutdown t.flow `All
with | Eio.Io (Eio.Net.E Connection_reset _, _) as ex ->
Log.info (fun f -> f "%a" Eio.Exn.pp ex);
This happens with non-TCP sockets. Reported by Patrick Ferris.
Instead of spawning a new flush fiber if one isn't running, just keep one around at all times. This is simpler, makes the traces cleaner, and will probably help with buffering later.
Sending each message in its own TCP packet isn't very efficient.
Note: The Prometheus metrics `messages_outbound_sent_total` and
`messages_outbound_dropped_total` have gone. They weren't very useful
and we no longer know the number of messages by the time the connection
is dropped (could report dropped bytes if needed though).
$ dune exec -- ./test-bin/echo/echo_bench.exe
echo_bench.exe: [INFO] rate = 18466.209374 # Before
echo_bench.exe: [INFO] rate = 59655.912455 # After
This switches capnp-rpc from Lwt to Eio. One particularly nice side effect of this is that
Service.return_lwthas gone, as there is no distinction now between concurrent and non-concurrent service methods.Notes:
capnp_rpctocapnp_rpc_protocolor something, leaving the short name free for the main library.This commit aims to keep the diff (relatively) small. Several more things are needed after this is merged. In particular:
Exitto finish.