perf: various improvements

Author: dignifiedquire

Cargo.lock

@@ -43,6 +43,8 @@ unused-async = "warn"
 
 
 [patch.crates-io]
+n0-watcher = { git = "https://github.com/n0-computer/n0-watcher", branch = "feat-get-ref" }
+
 iroh-quinn = { git = "https://github.com/n0-computer/quinn", branch = "main-iroh" }
 iroh-quinn-proto = { git = "https://github.com/n0-computer/quinn", branch = "main-iroh" }
 iroh-quinn-udp = { git = "https://github.com/n0-computer/quinn", branch = "main-iroh" }
@@ -52,3 +54,10 @@ netwatch = { git = "https://github.com/n0-computer/net-tools", branch = "main" }
 # iroh-quinn = { path = "../iroh-quinn/quinn" }
 # iroh-quinn-proto = { path = "../iroh-quinn/quinn-proto" }
 # iroh-quinn-udp = { path = "../iroh-quinn/quinn-udp" }
+
+
+[patch."https://github.com/n0-computer/quinn"]
+
+# iroh-quinn = { path = "../iroh-quinn/quinn" }
+# iroh-quinn-proto = { path = "../iroh-quinn/quinn-proto" }
+# iroh-quinn-udp = { path = "../iroh-quinn/quinn-udp" }

Cargo.toml

@@ -50,6 +50,7 @@ rustls = { version = "0.23.33", default-features = false, features = ["ring"] }
 serde = { version = "1.0.219", features = ["derive", "rc"] }
 smallvec = "1.11.1"
 strum = { version = "0.27", features = ["derive"] }
+tinyvec = { version = "1.10.0", features = ["alloc"] }
 tokio = { version = "1.44.1", features = [
     "io-util",
     "macros",

iroh/Cargo.toml

@@ -75,7 +75,10 @@ pub(crate) mod endpoint_map;
 pub(crate) mod mapped_addrs;
 pub(crate) mod transports;
 
-use mapped_addrs::{EndpointIdMappedAddr, MappedAddr};
+use self::{
+    mapped_addrs::{EndpointIdMappedAddr, MappedAddr},
+    transports::Addr,
+};
 
 pub use self::{endpoint_map::PathInfo, metrics::Metrics};
 
@@ -489,18 +492,18 @@ impl MagicSock {
 
     #[cfg_attr(windows, allow(dead_code))]
     fn normalized_local_addr(&self) -> io::Result<SocketAddr> {
-        let addrs = self.local_addrs_watch.clone().get();
+        let addrs = self.local_addrs_watch.get_ref();
 
         let mut ipv4_addr = None;
         for addr in addrs {
-            let Some(addr) = addr.into_socket_addr() else {
-                continue;
-            };
-            if addr.is_ipv6() {
-                return Ok(addr);
-            }
-            if addr.is_ipv4() && ipv4_addr.is_none() {
-                ipv4_addr.replace(addr);
+            match addr {
+                Addr::Ip(addr @ SocketAddr::V6(_)) => {
+                    return Ok(*addr);
+                }
+                Addr::Ip(addr @ SocketAddr::V4(_)) if ipv4_addr.is_none() => {
+                    ipv4_addr.replace(*addr);
+                }
+                _ => {}
             }
         }
         match ipv4_addr {
@@ -550,11 +553,12 @@ impl MagicSock {
 
         let mut quic_packets_total = 0;
 
-        for ((quinn_meta, buf), source_addr) in metas
-            .iter_mut()
-            .zip(bufs.iter_mut())
-            .zip(source_addrs.iter())
-        {
+        // zip is slow :(
+        for i in 0..metas.len() {
+            let quinn_meta = &mut metas[i];
+            let buf = &mut bufs[i];
+            let source_addr = &source_addrs[i];
+
             let mut buf_contains_quic_datagrams = false;
             let mut quic_datagram_count = 0;
             if quinn_meta.len > quinn_meta.stride {
@@ -578,11 +582,9 @@ impl MagicSock {
                 // relies on quinn::EndpointConfig::grease_quic_bit being set to `false`,
                 // which we do in Endpoint::bind.
                 if let Some((sender, sealed_box)) = disco::source_and_box(datagram) {
-                    trace!(src = ?source_addr, len = datagram.len(), "UDP recv: DISCO packet");
                     self.handle_disco_message(sender, sealed_box, source_addr);
                     datagram[0] = 0u8;
                 } else {
-                    trace!(src = ?source_addr, len = datagram.len(), "UDP recv: QUIC packet");
                     match source_addr {
                         transports::Addr::Ip(SocketAddr::V4(..)) => {
                             self.metrics

iroh/src/magicsock.rs

@@ -3,7 +3,7 @@ use std::{
     io::{self, IoSliceMut},
     net::{IpAddr, Ipv6Addr, SocketAddr, SocketAddrV6},
     pin::Pin,
-    sync::{Arc, atomic::AtomicUsize},
+    sync::Arc,
     task::{Context, Poll},
 };
 
@@ -34,7 +34,8 @@ pub(crate) struct Transports {
     ip: Vec<IpTransport>,
     relay: Vec<RelayTransport>,
 
-    poll_recv_counter: AtomicUsize,
+    poll_recv_counter: usize,
+    source_addrs: tinyvec::TinyVec<[Addr; 4]>, // cache for source addrs
 }
 
 #[cfg(not(wasm_browser))]
@@ -69,6 +70,7 @@ impl Transports {
             ip,
             relay,
             poll_recv_counter: Default::default(),
+            source_addrs: Default::default(),
         }
     }
 
@@ -84,11 +86,11 @@ impl Transports {
             return Poll::Pending;
         }
 
-        let mut source_addrs = vec![Addr::default(); metas.len()];
-        match self.inner_poll_recv(cx, bufs, metas, &mut source_addrs)? {
+        self.source_addrs.resize_with(metas.len(), Addr::default);
+        match self.inner_poll_recv(cx, bufs, metas)? {
             Poll::Pending | Poll::Ready(0) => Poll::Pending,
             Poll::Ready(n) => {
-                msock.process_datagrams(&mut bufs[..n], &mut metas[..n], &source_addrs[..n]);
+                msock.process_datagrams(&mut bufs[..n], &mut metas[..n], &self.source_addrs[..n]);
                 Poll::Ready(Ok(n))
             }
         }
@@ -100,13 +102,12 @@ impl Transports {
         cx: &mut Context,
         bufs: &mut [IoSliceMut<'_>],
         metas: &mut [quinn_udp::RecvMeta],
-        source_addrs: &mut [Addr],
     ) -> Poll<io::Result<usize>> {
         debug_assert_eq!(bufs.len(), metas.len(), "non matching bufs & metas");
 
         macro_rules! poll_transport {
             ($socket:expr) => {
-                match $socket.poll_recv(cx, bufs, metas, source_addrs)? {
+                match $socket.poll_recv(cx, bufs, metas, &mut self.source_addrs)? {
                     Poll::Pending | Poll::Ready(0) => {}
                     Poll::Ready(n) => {
                         return Poll::Ready(Ok(n));
@@ -117,9 +118,7 @@ impl Transports {
 
         // To improve fairness, every other call reverses the ordering of polling.
 
-        let counter = self
-            .poll_recv_counter
-            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+        let counter = self.poll_recv_counter.wrapping_add(1);
 
         if counter % 2 == 0 {
             #[cfg(not(wasm_browser))]