perf: enhance WireGuard performance

Cargo.toml

@@ -16,6 +16,6 @@ debug = true
 
 [patch.crates-io]
 rustls = { git = "https://github.com/XOR-op/rustls.delta.git", branch = "unofficial-rel-0.23" }
-smoltcp = { git = "https://github.com/XOR-op/smoltcp.git", branch = "resize-recv-buffer" }
+smoltcp = { git = "https://github.com/XOR-op/smoltcp.git", branch = "rcv-buf-pinned" }
 # Only used to bump x25519-dalek; will be removed once 0.6.1 is released
 boringtun = { git = "https://github.com/XOR-op/boringtun", branch = "master"}

boltconn/Cargo.toml

@@ -39,9 +39,10 @@ network-interface = "2.0.0"
 nix = { version = "0.29.0", features = ["user", "fs"] }
 rand = { version = "0.8.5", features = ["small_rng"] }
 regex = "1.7.0"
+sharded-slab = "0.1.7"
 socket2 = { version = "0.5.1", features = ["all"] }
 thiserror = "1.0.37"
-tokio = { version = "1.32.0", features = [
+tokio = { version = "1.40.0", features = [
     "rt",
     "rt-multi-thread",
     "net",

boltconn/src/common/mod.rs

@@ -19,8 +19,8 @@ pub mod client_hello;
 pub mod duplex_chan;
 pub mod evictable_vec;
 pub mod host_matcher;
-pub mod id_gen;
 mod sync;
+pub mod utils;
 
 pub use sync::{local_async_run, AbortCanary};
 
@@ -45,6 +45,7 @@ impl StreamOutboundTrait for tokio::net::windows::named_pipe::NamedPipeServer {}
 impl StreamOutboundTrait for tokio::net::UnixStream {}
 
 pub const MAX_PKT_SIZE: usize = 65576;
+pub const MAX_UDP_PKT_SIZE: usize = 1518;
 
 pub async fn read_to_bytes_mut(
     buf: &mut BytesMut,

boltconn/src/common/utils.rs

@@ -0,0 +1,53 @@
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::{Duration, Instant};
+
+#[derive(Default, Debug)]
+pub struct IdGenerator(AtomicU64);
+
+impl IdGenerator {
+    pub fn get(&self) -> u64 {
+        self.0.fetch_add(1, Ordering::Relaxed)
+    }
+}
+
+#[allow(unused)]
+pub struct TputMeasurement {
+    inner: std::sync::Mutex<TputMeasurementInner>,
+}
+
+struct TputMeasurementInner {
+    pub accum_bytes: u64,
+    pub last_time: Instant,
+    pub interval: Duration,
+}
+
+impl TputMeasurement {
+    pub fn new(interval: Duration) -> Self {
+        Self {
+            inner: std::sync::Mutex::new(TputMeasurementInner {
+                accum_bytes: 0,
+                last_time: Instant::now(),
+                interval,
+            }),
+        }
+    }
+
+    pub fn update(&self, bytes: usize) -> Option<f64> {
+        let mut inner = self.inner.lock().unwrap();
+        inner.accum_bytes += bytes as u64;
+        let now = Instant::now();
+        let elapsed = now - inner.last_time;
+        if elapsed >= inner.interval {
+            let tput = inner.accum_bytes as f64 / elapsed.as_secs_f64();
+            inner.accum_bytes = 0;
+            inner.last_time = now;
+            Some(tput)
+        } else {
+            None
+        }
+    }
+
+    pub fn update_to_mbps(&self, bytes: usize) -> Option<f64> {
+        self.update(bytes).map(|tput| tput * 8.0 / 1_000_000.0)
+    }
+}