When a stream is opened we are searching for the best (currently) available channel.
The stream will then be keept on that channel.
Adjusted the rest of the algorithms that they now respect this rule.
improved a HashMap for Pids to be based on a Vec. Also using this for Sid -> Cid relation which is more performance critical
WARN: our current send()? error handling allows it for some close_stream messages to get lost.
- now last digit version is compatible 0.6.0 will connect to 0.6.1
- the TCP DATA Frames no longer contain START field, as it's not needed
- the TCP OPENSTREAM Frames will now contain the BANDWIDTH field
- MID is not Protocol internal
Update network
- update API with Bandwidth
Update veloren
- introduce better runtime and `async` things that are IO bound.
- Remove `uvth` and instead use `tokio::runtime::Runtime::spawn_blocking`
- remove futures_execute from client and server use tokio::runtime::Runtime instead
- give threads a Name
- a channel was stale and wasn't shut down propertly AS WELL AS
- the msg ingoing pipe was bounded, so it could fill up
To mitigate this we
a) unbounded the pipe
b) stoped spam the log in no channel case
c) instead of ever reaching "no channel" state we actually shutdown participant
d) when send_mgr is closed it will no longer be able to SEND on streams
- completly switch to Bytes, even in api. speed up TCP by fak 2
- improve benchmarks
- speed up mpsc metrics
- gracefully handle shutdown by interpreting Ok(0) as tokio::tcpstream closed now.
- fix hotloop in participants by adding `Some(n)` to fix endless handing.
- fix closing bug by closing streams after `recv_mgr` is shutdown even if now shutdown is triggered locally.
- fix prometheus
- no longer throw when a `Stream` is dropped while participant still receives a msg for it.
- fix the bandwith handling, TCP network send speed is up to 1.5GiB/s while recv is 150MiB/s
- add documentation
- tmp require rt-multi-threaded in client for tokio, to not fail cargo check
this is prob stable, i tested over 1 hour.
after that some optimisations in priomgr.
and impl. propper bandwith.
Speed is up to 2GB/s write and 150MB/s recv on a single core
sync add documentation
- better logging in network
- we now notify the send of what happened in recv in participant.
- works with veloren master servers
- works in singleplayer, using a actual mid.
- add `mpsc` in whole stack incl tests
- speed up internal read/write with `Bytes` crate
- use `prometheus-hyper` for metrics
- use a metrics cache
- Implementing a async non-io protocol crate
a) no tokio / no channels
b) I/O is based on abstraction Sink/Drain
c) different Protocols can have a different Drain Type
This allow MPSC to send its content without splitting up messages at all!
It allows UDP to have internal extra frames to care for security
It allows better abstraction for tests
Allows benchmarks on the mpsc variant
Custom Handshakes to allow sth like Quic protocol easily
- reduce the participant managers to 4: channel creations, send, recv and shutdown.
keeping the `mut data` in one manager removes the need for all RwLocks.
reducing complexity and parallel access problems
- more strategic participant shutdown. first send. then wait for remote side to notice recv stop, then remote side will stop send, then local side can stop recv.
- metrics are internally abstracted to fit protocol and network layer
- in this commit network/protocol tests work and network tests work someway, veloren compiles but does not work
- handshake compatible to async_std
switch to `tokio` and `async_channel` crate.
I wanted to do tokio first, but it doesnt feature Sender::close(), thus i included async_channel
Got rid of `futures` and only need `futures_core` and `futures_util`.
Tokio does not support `Stream` and `StreamExt` so for now i need to use `tokio-stream`, i think this will go in `std` in the future
Created `b2b_close_stream_opened_sender_r` as the shutdown procedure does not need a copy of a Sender, it just need to stop it.
Various adjustments, e.g. for `select!` which now requieres a `&mut` for oneshots.
Future things to do:
- Use some better signalling than oneshot<()> in some cases.
- Use a Watch for the Prio propergation (impl. it ofc)
- Use Bounded Channels in order to improve performance
- adjust tests coding
bring tests to work
But i have the feeling that maybe something with the channel is wrong which leads to this behavior (or maybe did i made a copy somewhere, though i dobt this).
Again, not sure if this is a fix, but i think it doesn't hurt
There is a rare bug that recently got triggered more often with the release of xMAC94x/netfixA
if the bug triggeres, a Participant never gets cleaned up gracefully.
Reason:
When `participant_shutdown_mgr` was called it stopped all managers at once.
Especially stream_close_mgr and send_mgr.
The problem with stream_close_mgr is, it's responsible for gracefully flushing streams when the Participant is dropped locally.
So when it was interupted self.streams where no longer flushed gracefully.
The next problem was with send_mgr.
It is triggering the PrioManager, and the PrioManager is responsible for notifying once a stream is completly flushed.
This lead to the problem, that a stream flush could be requested, but was actually never executed (as send_mgr was already down).
Solution:
1. when stream_close_mgr is stopped it MUST flush all remaining streams
2. wait for stream_close_mgr to finish before shutting down the send_mgr
3. no longer delete streams when closing the API (this also wasn't tracked in metrics so far)
Additionally i added a dependency, so that the network/examples compile again, fixed some spelling.
I created a `delete_stream` fn that basically just moved the code over.
- first remove participant AND channel in same metric. this caused a matrix full of 0 values which bloated alot.
- then did the cid cache to be lazy loading to no longer generate that much 0 values
- possible would also be no longer keeping metrics for INIT, HANDSHAKE and PARTICIPANTID
- When Participant A was closed by remote side. Then a `disconnect` on `A`
shall return Ok() (instead of ParticipantDisconected) IF:
A was already flushed and no data needs to be sended any more.
so a `disconnect` doesnt differ if the other side initiated the disconnect before or not. it tries to clean things up and returns Ok(()) if both sides agree
- It was possible for a end_receiver to be triggered in the moment while a frame was started by not finished.
This removed bytes from the stream with them getting lost. this almost ever was followed by a RAW frame as the TCP stream was now invalid.
The TCP stream was then detected by participant or caused one or multiple failures
- introduces some simplifications, removed a macro, reuse code
- replace RwLock by Mutex if it's only accessed for insert/delete
- use RwLock<HashMap<Mutex>> pattern otherwise in order to allow concurrent `.read()`
- fixed a deadlock O.o
Till now, we just dropped the TCP connection and registered this as a clean shutdown.
The prodocol reader intereted this and send a Frame::Shutdown frame to it's local processor.
This is ofc wrong.
So now the protocol reader will detect a Frame::Shutdown frame and send it over. if the Tcp connection gets closed it will return an Error up.
The processor will then pick up this error and request a unclear shutdown and notifies the user.
Also when doing a clean shutdown we are sending a Frame::Shutdown now to the remote side to trigger this behavior.
Before we wrongly added the feature of only using a `select` in channel. This is WRONG,
as it could mean that the write maybe fails, but the read still had some Frames buffered which then get dropped.
Its fixed now by the clean shutdown mechanims defined before.
Also when a channel is closed now inside a participant we are closing the whole participant as a protection.
However, we must not close the recv channel as the `handle_frames_mgr` might still be working on them, so we only stop writing/sending.
Debugging this also let me introduce some smaller fixes:
- PID in tests are now 0 and 1+1*64+1*64*64+... this makes the traces appear as AAAAAA and BBBBBB instead of ABAAAA and ACAAAA
- veloren client now better seperates between clean shutdown and unclear shutdown.
- added a new type: C2pFrame for `(cid, Result<Frame, ()>)`
- wrong frames inside the handshare are not counted in metrics
-
added a span for disconnecting on the gameserver side. also added more debug! tracing there
Just keeping a trace! all 10000ms active to have a keep alive feeling.
Participant will now handle a close in the `create_channel_mgr` rather then the `send` fn. Its the better place, which makes a HashMap better for delete lookup
Since tcp_read now broke but tcp_write didn't and the Participant wasnt updated till both were broke, we changed CHANNEL tcp_read and tcp_write in protocols to be a `select` rather than a `join`
However only do this in the CHANNEL, but in the HANDSHAKE. it fails if you try to. Also the handshake will take care of any failed read or write manually and will handle a clear teardown in this case.
increase the slowlorris timeout. for some reason it seems to trigger alot more often since commit: `75c1d440` but i have no idea why.
My guess would be that the initial sync now sends way more data which slows down TCP to be > 10ms and trigger. Note: the fix might cause small lags when slow people try to connect to the server
The remote site should see it connect, be open for 1 single stream and read the message before it's notified that the participant is closed actually.
This caused the faulure in one of our API tests (in lib, with client and server). Where it was possible that all messages were send and one side was dropped before the other side asked for the opened stream
Also introduce better error detection in participant(and scheduler) by removing the std_async::Result and intruduce `Result<(),ParticipantError>` instead
- we had the problem that Participants couldn't shutdown them self, only by scheduler, which was controlled by api.
it's needed e.g. to handle the Schudown Frame
- my initial solution did a full shutdown, which was a problem if in parallel a 2nd shutdown was requested, no possibility of getting the error
- new solution will only deactivate Participant and Stream. and then still functions correctly, till the api closes the participant and calls the scheduler which then calls the bparticipant again
- i experimented with a Mutex<oneshot> or 2 and a `select` but it didn't prove that well
- also adjusted the Error messages to now either Disconnected when gracefully shutdown or ProtocolFailed when some msg couldn't be delivered
(note later might not be 100% returned correctly yet)
- API behavior switched!
- the `Network` no longer holds a copy of participant, thus if the return of `connect` (before `Arc<Participant>, now `Participant`) got dropped, the `Participant::Drop` is triggered!
- you can close a Participant async via `Particiant::disconnect()`, no more need to know the network at this point
- the `Network::Drop` will check and drop not yet disconnected Participants.
- you can compare Participants via PartialEq, if they are true they point to the same endpoint (it checks remote_pid)
- Note: multiple Participants are only supported in theory, wont work yet
Additionally:
- fix some `debug!`
- veloren-client will now drop the participant gracefully on shutdown
- rename `error` to `debug` when 2 times Bparticipant shutdown is called, as it is to be expected in a async runtime
- swap out std::mpsc with crossbeam-channel in networking crate
- remove log spam by only logging when populating a new cache entry and not on every get
- switched the behavior of prio to keep the order of a stream, EVEN if messages of different length are used!
This needs to be addresses to fulfill PROMISES_ORDERED
I think we need to make thoughs if PRIO needs to handle this, or we add additional meta info and handle it on receiver site
- added PartialEq to StreamError for test purposes (only yet!)
- removed async_recv example as it's no longer for any use.
It was created before the COMPLETE REWRITE in order to verify that my own async interface on top of mio works.
However it's now guaranteed by async-std and futures. no need for a special test
- remove uvth from dependencies and replace it with a `FnOnce`
- fix ALL clippy (network) lints
- basic fix for a channel drop scenario:
TODO: this needs some further fixes
up to know only destruction of participant by api was covered correctly.
we had an issue when the underlying channels got dropped. So now we have a participant without channels.
We need to buffer the requests and try to reopen a channel ASAP!
If no channel could be reopened we need to close the Participant, while
a) leaving the BParticipant in takt, knowing that it only waits for a propper close by scheduler
b) close the BParticipant gracefully. Notifying the scheduler to remove its stuff (either scheduler schould detect a stopped BParticipant or BParticipant will send Scheduler it's own destruction, and then Scheduler just does the same like when API forces a close)
Keep the Participant alive and wait for the api to acces BParticipant to notice it's closed and then wait for a disconnect which isn't doing anything as it was already cleaned up in the background
- this bug was initially called imbris bug, as it happened on his runners and i couldn't reproduce it locally at fist :)
- When in a Handshake a seperate mpsc::Channel was created for (Cid, Frame) transport
however the protocol could already catch non handshake data any more and push in into this
mpsc::Channel.
Then this channel got dropped and a fresh one was created for the network::Channel.
These droped Frames are ofc a BUG!
I tried multiple things to solve this:
- dont create a new mpsc::Channel, but instead bind it to the Protocol itself and always use 1.
This would work theoretically, but in bParticipant side we are using 1 mpsc::Channel<(Cid, Frame)>
to handle ALL the network::channel.
If now ever Protocol would have it's own, and with that every network::Channel had it's own it would no longer work out
Bad Idea...
- using the first method but creating the mpsc::Channel inside the scheduler instead protocol neither works, as the
scheduler doesnt know the remote_pid yet
- i dont want a hack to say the protocol only listen to 2 messages and then stop no matter what
So i switched over to the simply method now:
- Do everything like before with 2 mpsc::Channels
- after the handshake. close the receiver and listen for all remaining (cid, frame) combinations
- when starting the channel, reapply them to the new sender/listener combination
- added tracing
- switched Protocol RwLock to Mutex, as it's only ever 1
- Additionally changed the layout and introduces the c2w_frame_s and w2s_cid_frame_s name schema
- Fixed a bug in scheduler which WOULD cause a DEADLOCK if handshake would fail
- fixd a but in api_send_send_main, i need to store the stream_p otherwise it's immeadiatly closed and a stream_a.send() isn't guaranteed
- add extra test to verify that a send message is received even if the Stream is already closed
- changed OutGoing to Outgoing
- fixed a bug that `metrics.tick()` was never called
- removed 2 unused nightly features and added `deny_code`
fix async_recv and double block_on panic on Network::drop and participant::drop
include Cargo.lock from all examples
Found a bug on imbris runners with doc tests of `stream::send` and `stream::recv`
As neither a backtrace, nor tracing on runners in the doc tests seems to help, i disable them and add them as unit tests
however i need to coordinate the prio adjustments in scheduler from now on, so that ParticipantA doesn't get all the network bandwith and ParticipantB nothing
- https://github.com/tikv/rust-prometheus/issues/321
- split up channel into a hanshake part and channel part.
The handshake part is non endless and ends when its either done or aborted.
If its okay i will send a request to the BParticipant which then opens a channel on the existing TCP or UDP connection.
this streamlines the command chain alot. also the channel is almost empty now, thinking about removing it completly.
isnt perfect, as shutdown and udp doesnt work yet
- make PID to print as Base64
- replace rouille with tiny_http
