According to RUSTSEC-2023-0052 we need to upgrade rustls to 0.21
to get a fix for the issue which may or may not affect Veloren
since it's about client certificates but with the absence of a
PoC it seems like a good idea to upgrade anyway, just to be sure.
webpki has gone unmaintained (which rustls 0.20 depends on),
starting with rustls 0.21 it depends on rustls-webpki which
contains a fix for the issue. Since quinn also depends on
rustls 0.20 in 0.8 and 0.9 versions, we needed to upgrade it to
0.10 so that it depends on rustls 0.21 which we now use.
and because of that we decided to set the recv buffer to 512kB via linux kernel parameters.
We didn't had any problems on the official server so we wanted to bake that into code.
There is a calculator https://www.switch.ch/network/tools/tcp_throughput/?mss=1460&rtt=80&loss=1e-06&bw=10&rtt2=500&win=1024&Calculate=Calculate
that said for 10Mbits/ speed 16kB recv buffer and 500ms we only can expect effective 0.24 Mbits of actuall data speed.
Nothing mentioned increasing the send buffer, but it prob cant hurt from a server side or ?
cat /proc/sys/net/core/wmem_default showed 212992
cat /proc/sys/net/core/rmem_default showed 212992
sysctl -w net.ipv4.tcp_wmem="4096 524288 4194304" was beeing used
sysctl -w net.ipv4.tcp_wmem="4096 16384 4194304" was the default
First we though about hashing it together with the username, but we dont have the username. hope that removing 50% of the information is enough to identify 2 addresses as equal or different without knowing the exact address
Because the old case didn't account for the Scheduler got dropped but Participant keept around.
Shutdown behavior is quite easy now: bParticipant sends a oneshot, when that hits we stop. also when Participant gets droped we still stop as there woul dbe no sense in continue running the report_mgr ...
This will ask the bparticipant for a list of all channels and their respective connection arguments.
With that one could prob reach the remote side.
The data is gathered by scheduler (or channel for the listener code).
It requeres some read logs so we shouldn't abuse that function call.
in bparticipant we have a new manager that also properly shuts down as the Participant holds the sender to the respective receiver.
The sender is always dropped. inside the Mutex via disconnect and outside via Drop (we need 2 Options as otherwise we would create a runtime inside async context implicitly o.O )
(also i didn't liked the alternative by just overwriting the sender with a fake one, i want a propper Option that can be taken)
The code might also come handy in the future when we implement a auto-reconnect feature in the bparticipant.
Note: maybe upgrading tokio helps against the connection issues, there is some mio related stuff: https://github.com/tokio-rs/tokio/releases/tag/tokio-1.18.0
Note: quick also still has the old read, but its more complicated, thus i only want to change it if its tested
We calculate a max bandwidth, but due to framing the packages, i think its possible for the current_bytes to get a few bytes over the actuall planed bandwidth. which prob isn't that bad/unfair
for that we need to create a socket2 socket and set_only_v6 to true.
We need to privude set_nonblocking for tokio to work and we need to set reuseaddr for it to work on linux systems
cannot update the following dependencies:
- vek: Sharps SIMD isnt upstream
- tracing-subscriber: MakeWriter was adjusted and i was to lazy to fiddle with lifetimes,
- refinery, rustsql: we have a custom refinery version which is incompatible with newer rustsql
- equi + egui_winit + egui_wgpu_backend: i tried it in this commit but it turned out that they dependo n wgpu which we cant update
- wgpu: cant update due new version doesnt support DX11
Got quinn updated which now require some dependencies to be explicit.
Increase the timeouts from 1000 -> 3000 ms if there are some internal messages send (e.g. network_a closed, send on stream_a)
Increase the timeouts from 1000 -> 5000 ms if there is actuall networking involed (e.g. stream_a send, stream_b recv)
