Repository navigation
TCP/IP + Basic async runtime. - #6
Merged
Merged
Conversation
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## main #6 +/- ##
==========================================
+ Coverage 88.37% 89.60% +1.23%
==========================================
Files 44 101 +57
Lines 3767 22433 +18666
Branches 3767 22433 +18666
==========================================
+ Hits 3329 20101 +16772
- Misses 375 2138 +1763
- Partials 63 194 +131 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
…errors from fill queue.
… UDP/IPv4/IPv6/TCP.
…ypes for ethernet.
… net::wire::{ndp,icmpv4,icmpv6}.
…ekerless our futures.
neli's typed attribute deserialization misparses ethtool CHANNELS_GET responses that mix nested (Header) and flat (count) attributes at the same level. Switch to raw netlink socket I/O for the response and parse NLA attributes manually. Also treat 0xFFFFFFFF as "not applicable" (ethtool's encoding for n/a channels, common on virtio-net/cloud VMs).
The ethtool command ID was wrong: we used 4 (LINKMODES_GET) instead of 17 (CHANNELS_GET). The attribute enum was also misordered — OtherMax and OtherCount were placed at the end instead of their correct positions (4 and 8), shifting RxCount/TxCount/CombinedMax/CombinedCount by 2.
This was becoming a maintenance burden and at the end of the day wasn't especially viable as a operating mode the new rt will replace.
Design spec and implementation plan for making the event loop's TX completion drain non-blocking to fix multi-queue scaling regression (342K vs 544K req/s, 6.4x per-core efficiency drop).
Replace the synchronous completion drain in LocalRuntime::run() with an in-flight TX counter. Completions are collected opportunistically at the top of each iteration rather than blocking until all arrive. This removes the per-iteration kernel round-trip that was throttling multi-queue throughput. Key changes: - Add in_flight_tx counter tracking frames in kernel TX pipeline - Drain completions non-blocking at loop start (Phase 1) - Send tx_return non-blocking with one retry on WouldBlock - Extract recycle_rx_return() helper (fill queue first, overflow to free_frames) - Fix pre-existing invariant gap: use <= to account for FragmentReader frames New invariant: free_frames + in_flight_tx + tx_return <= expected_total (replaces strict free_frames == expected_free_frames) Spec: docs/superpowers/specs/2026-03-17-async-tx-completion-design.md
socket.send() only writes descriptors to the TX ring — the kernel needs a sendto() syscall (via maybe_wake) to start DMA processing. The async completion drain removed the synchronous completion loop which had been implicitly providing this kick. Without it, frames sat in the TX ring and never transmitted.
NeighborHandler is only accessed single-threaded via Rc within LocalRuntime, but was using DashMap (a concurrent hashmap) for the neighbor table. Perf profiling showed 22.7% of CPU in lookup_or_resolve and another 15.4% in DashMap's atomic CAS operations (cas8_acq/cas8_rel) — a total of ~38% CPU wasted on unnecessary synchronization. Replace DashMap with FxHashMap wrapped in UnsafeCell for interior mutability (same pattern as PmtuCache and TcpHandler). This eliminates all atomic operations, shard locking, and replaces SipHash with FxHash for lookups. Changes: - handler.rs: DashMap → UnsafeCell<FxHashMap>, add table() helper - arp.rs: &DashMap param → &mut FxHashMap - ndp.rs: &DashMap param → &mut FxHashMap (4 functions) - Cargo.toml: remove dashmap dependency
…ghbor lookup Perf profiling showed lookup_or_resolve consuming 33% of CPU — called on every TCP segment in poll_send, doing a FxHashMap probe each time even though the destination MAC never changes for an established connection. Add dst_mac: Option<MacAddress> to Tcb. On first successful neighbor resolution, the MAC is cached. All subsequent segments on that connection use the cached value directly, bypassing the hash table entirely. 5 call sites in transmit.rs updated with cache-first pattern: - Data segment send loop (hot path) - Delayed ACK fallback - Zero-window probe - Linger RST - FIN transmission
The previous MAC caching commit only updated transmit.rs (poll_send), but timers.rs has 4 additional lookup_or_resolve call sites that were bypassing the cache. More critically, listen.rs was initializing dst_mac: None even though the inbound path already knows the remote MAC from the incoming Ethernet frame header. Fixes: - listen.rs: initialize dst_mac from incoming frame's src_mac - timers.rs: add cache-first pattern to all 4 lookup_or_resolve sites (delayed ACK, keep-alive, SACK recovery, RTO retransmit)
Add mpsc channel mesh so multi-queue workers propagate ARP/NDP neighbor updates to peer queues. Each LocalRuntime drains its receiver at the top of the event loop before packet processing.
… suppress private_interfaces warnings
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This implements a partial but robust network stack including: