It seems the root cause is if the flag set at the inbound pipe reader, it is a race condition and freedom outbound can possibly do splice at the same time with inbound xtls writer.
Now we set the flag at the earliest and always do splice at the next buffer cycle.
- Vision now use traffic states to capture two-way info about a connection
- XTLS is de-couple with Vision, it only read traffic states to switch to direct copy mode
- fix a edge case error when Vision unpadding read 5 command bytes
- Add outbound name
- Add outbound conn in ctx
- Refactor splice: it can be turn on from all inbounds and outbounds
- Refactor splice: Add splice copy to vless inbound
- Fix http error test
- Add freedom splice toggle via env var
- Populate outbound obj in context
- Use CanSpliceCopy to mark a connection
- Turn off splice by default
* Vision server allow multiple blocks of padding
* Fix Vision client to support multiple possible padding blocks
* Vision padding upgrade
- Now we have two types of padding: long (pad to 900-1400) and traditional (0-256)
- Long padding is applied to tls handshakes and first (empty) packet
- Traditional padding is applied to all beginning (7) packets of the connection (counted two-way)
- Since receiver changed its way to unpad buffer in fd6973b3c6, we can freely extend padding packet length easily in the future
- Simplify code
* Adjust receiver withinPaddingBuffers
Now default withinPaddingBuffers = true to give it a chance to do unpadding
* Fix magic numbers for Vision
Thanks @H1JK
Thanks @RPRX for guidance
This only affects the Vision client for protocols expecting server to send data first.
The change is compatible with existing version of Vision server.
In the past, when user open xtls vision on the server side, plain vless+tls can connect.
Pure tls is known to have certain tls in tls characters.
Now server need to specify "xtls-rprx-vision,none" for it be able usable on the same port.
Before this change, Vision client need a pure inbound like socks or http.
After this change, it will support any inbound.
This is useful in traffic forwarder use case inside China.
* Add XTLS RPRX's Vision
* Add helpful warning when security is wrong
* Add XTLS padding (draft)
* Fix number of packet to filter
* Xtls padding version 1.0 and unpadding logic
* DialSystem for Quic
DialSystem() is needed in case of Android client,
where the raw conn is protected for vpn service
* Fix client dialer log
Log such as:
tunneling request to tcp:www.google.com:80 via tcp:x.x.x.x:443
the second "tcp" is misleading when using mKcp or quic transport
Remove the second "tcp" and add the correct logging for transport dialer:
- transport/internet/tcp: dialing TCP to tcp:x.x.x.x:443
- transport/internet/quic: dialing quic to udp:x.x.x.x:443
* Quic new stream allocation mode
Currently this is how Quic works: client muxing all tcp and udp traffic through a single session, when there are more than 32 running streams in the session,
the next stream request will fail and open with a new session (port). Imagine lineup the session from left to right:
|
| |
| | |
As the streams finishes, we still open stream from the left, original session. So the base session will always be there and new sessions on the right come and go.
However, either due to QOS or bugs in Quic implementation, the traffic "wear out" the base session. It will become slower and in the end not receiving any data from server side.
I couldn't figure out a solution for this problem at the moment, as a workaround:
| |
| | |
| | |
I came up with this new stream allocation mode, that it will never open new streams in the old sessions, but only from current or new session from right.
The keeplive config is turned off from server and client side. This way old sessions will natually close and new sessions keep generating.
Note the frequency of new session is still controlled by the server side. Server can assign a large max stream limit. In this case the new allocation mode will be similar to the current mode.
* Increase some tls test timeout
* Fix TestUserValidator
* Change all tests to VMessAEAD
Old VMess MD5 tests will be rejected and fail in 2022
* Chore: auto format code
