v1.6.4: fix Full-mode L7 muxer not batching ops (#231)

The batch-build loop blocked on a 30 ms timeout for the first message,
then drained whatever else was in the channel via try_recv() and fired
the batch. Under any non-bursty workload, the channel queue was always
empty by the time the first op woke us up — so every "batch" had
exactly one op, defeating the entire batching premise. Reporter (w0l4i)
saw `batch: 1 ops → ..., rtt=6.3 s` repeating in logs even under high
concurrency.

Fix: after the first op lands, hold the buffer open for an 8 ms
coalescing window. Concurrent ops (parallel fetches, HTTP/2 stream
openings, etc.) now accumulate into the same batch. 8 ms is rounding
error against the 2–7 s Apps Script RTT we're amortizing, and restores
the multi-op-per-batch behavior the rest of the code already supports
(MAX_BATCH_OPS=50, MAX_BATCH_PAYLOAD_BYTES=4 MiB).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Cargo.lock

@@ -2186,7 +2186,7 @@ dependencies = [
 
 [[package]]
 name = "mhrv-rs"
-version = "1.6.3"
+version = "1.6.4"
 dependencies = [
  "base64 0.22.1",
  "bytes",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "mhrv-rs"
-version = "1.6.3"
+version = "1.6.4"
 edition = "2021"
 description = "Rust port of MasterHttpRelayVPN -- DPI bypass via Google Apps Script relay with domain fronting"
 license = "MIT"

android/app/build.gradle.kts

@@ -14,8 +14,8 @@ android {
         applicationId = "com.therealaleph.mhrv"
         minSdk = 24 // Android 7.0 — covers 99%+ of live devices.
         targetSdk = 34
-        versionCode = 142
-        versionName = "1.6.3"
+        versionCode = 143
+        versionName = "1.6.4"
 
         // Ship all four mainstream Android ABIs:
         //   - arm64-v8a      — 95%+ of real-world Android phones since 2019

docs/changelog/v1.6.4.md

@@ -0,0 +1,4 @@
+<!-- see docs/changelog/v1.1.0.md for the file format: Persian, then `---`, then English. -->
+• رفع باگ "L7 multiplexer در Full mode batch نمی‌کنه" ([#231](https://github.com/therealaleph/MasterHttpRelayVPN-RUST/issues/231)): در حالت Full، انتظار می‌رفت که چند op به یک batch HTTP request به Apps Script ترکیب بشن (`batch: 5 ops` یا `batch: 10 ops`)، ولی log نشون می‌داد همیشه `batch: 1 ops` — یعنی هر op جدا یه round-trip Apps Script می‌گرفت (که هر کدوم 2 تا 7 ثانیه طول می‌کشن). علت: loop دریافت پیام بلافاصله بعد از اولین message با `try_recv()` (non-blocking) صف رو drain می‌کرد، بدون pause برای جمع‌آوری بقیه ops. **Fix:** بعد از اولین op، یه پنجرهٔ ۸ میلی‌ثانیه‌ای باز می‌مونه تا opهای بعدی (مثل parallel fetches، HTTP/2 streams) همون batch رو پر کنن. ۸ms در مقابل ~۲ تا ۷ ثانیه RTT Apps Script اصلاً ناچیزه ولی efficiency batching رو برمی‌گردونه. ریپورت شده توسط w0l4i با log واضح
+---
+• Fix "L7 multiplexer not batching in Full mode" bug ([#231](https://github.com/therealaleph/MasterHttpRelayVPN-RUST/issues/231)): in `full` mode, multiple ops should coalesce into a single batched HTTP request to Apps Script (`batch: 5 ops` or `batch: 10 ops`), but logs showed `batch: 1 ops` consistently — each op got its own Apps Script round-trip (2-7 s each). Cause: the receive loop drained the channel via `try_recv()` (non-blocking) immediately after the first message arrived, with no window to let concurrent ops accumulate. **Fix:** after the first op lands, hold the buffer open for an 8 ms coalescing window so concurrent ops (parallel fetches, HTTP/2 stream openings, etc.) land in the same batch. 8 ms is rounding error against the ~2-7 s Apps Script RTT but restores the entire batching premise. Reported by w0l4i with a clean log snippet

src/tunnel_client.rs

@@ -55,6 +55,16 @@ const REPLY_TIMEOUT: Duration = Duration::from_secs(35);
 /// connect saves one Apps Script round-trip per new flow.
 const CLIENT_FIRST_DATA_WAIT: Duration = Duration::from_millis(50);
 
+/// How long the muxer holds open the batch buffer after the first op
+/// arrives, waiting for more ops to coalesce. Issue #231 — the previous
+/// implementation drained `try_recv()` *immediately* after the first
+/// message landed, so under any non-bursty workload every batch held
+/// exactly one op (defeating the entire batching premise). 8 ms is small
+/// vs the ~2-7 s Apps Script round-trip the batch is amortizing, but
+/// long enough that concurrent HTTP/2 stream openings, parallel fetches,
+/// or any other burst lands in the same batch.
+const BATCH_COALESCE_WINDOW: Duration = Duration::from_millis(8);
+
 /// Structured error code the tunnel-node returns when it doesn't know the
 /// op (version mismatch). Must match `tunnel-node/src/main.rs`.
 const CODE_UNSUPPORTED_OP: &str = "UNSUPPORTED_OP";
@@ -319,13 +329,29 @@ async fn mux_loop(mut rx: mpsc::Receiver<MuxMsg>, fronter: Arc<DomainFronter>) {
 
     loop {
         let mut msgs = Vec::new();
-        match tokio::time::timeout(Duration::from_millis(30), rx.recv()).await {
-            Ok(Some(msg)) => msgs.push(msg),
-            Ok(None) => break,
-            Err(_) => continue,
+        // Block on the first message — no point waking up to find an empty
+        // queue. Once the first op lands, we hold open BATCH_COALESCE_WINDOW
+        // so concurrent ops (parallel fetches, HTTP/2 stream openings, etc.)
+        // land in the same batch instead of getting a fresh round-trip each.
+        match rx.recv().await {
+            Some(msg) => msgs.push(msg),
+            None => break,
         }
-        while let Ok(msg) = rx.try_recv() {
-            msgs.push(msg);
+        let deadline = tokio::time::Instant::now() + BATCH_COALESCE_WINDOW;
+        loop {
+            // Drain anything that's already queued without waiting.
+            while let Ok(msg) = rx.try_recv() {
+                msgs.push(msg);
+            }
+            let now = tokio::time::Instant::now();
+            if now >= deadline {
+                break;
+            }
+            match tokio::time::timeout(deadline - now, rx.recv()).await {
+                Ok(Some(msg)) => msgs.push(msg),
+                Ok(None) => return,
+                Err(_) => break,
+            }
         }
 
         // Split: plain connects go parallel, data-bearing ops get batched.