perf(quic): lock-free hot paths — ThreadLocal Cipher, AtomicReference close, @Volatile getters

Audit of blocking and synchronized code in the QUIC module surfaced four
hot-path wins, all verified against :quic:jvmTest.

- PlatformCrypto: header-protection AES-ECB now uses a per-thread cached
  Cipher. Previously every inbound and outbound packet paid for
  Cipher.getInstance("AES/ECB/NoPadding") provider lookup. ThreadLocal is
  safe because the call is stateless — every invocation re-init's with the
  caller-supplied key.
- JdkCertificateValidator: gate the SAN-side InetAddress.getByName behind
  looksLikeIpLiteral so a malformed cert with a hostname in a type 7 SAN
  cannot trigger DNS resolution on the TLS validation path.
- QuicConnectionDriver.close: replace synchronized(this) double-checked
  init with AtomicReference<Job?> + compareAndSet on a CoroutineStart.LAZY
  job. Lock-free, removes a synchronized from commonMain, preserves the
  original "first caller wins, second awaits same Job" contract.
- SendBuffer: mark _nextOffset, nextSendOffset, _finPending, _finSent,
  _finAcked @Volatile and drop synchronized from their single-field
  getters (nextOffset, sentOffset, finPending, finSent, finAcked). The
  compound-formula readableBytes getter still synchronizes.

https://claude.ai/code/session_01CXTjnuHKCNXDmpfyKxgG3V

quic/src/commonMain/kotlin/com/vitorpamplona/quic/connection/QuicConnectionDriver.kt

@@ -31,6 +31,8 @@ import kotlinx.coroutines.joinAll
 import kotlinx.coroutines.launch
 import kotlinx.coroutines.sync.withLock
 import kotlinx.coroutines.withTimeoutOrNull
+import kotlin.concurrent.atomics.AtomicReference
+import kotlin.concurrent.atomics.ExperimentalAtomicApi
 
 /**
  * Owns the UDP socket and runs the read + send loops for a [QuicConnection].
@@ -48,6 +50,7 @@ import kotlinx.coroutines.withTimeoutOrNull
  * and [QuicConnection.openBidiStream]/[com.vitorpamplona.quic.stream.SendBuffer.enqueue])
  * call [wakeup] to nudge the send loop.
  */
+@OptIn(ExperimentalAtomicApi::class)
 class QuicConnectionDriver(
     val connection: QuicConnection,
     private val socket: UdpSocket,
@@ -94,17 +97,20 @@ class QuicConnectionDriver(
      * to poll `connection.status == CLOSED` and trust that the rest of
      * the cleanup eventually settled.
      */
-    internal val closeTeardownJob: Job? get() = closeJob
+    internal val closeTeardownJob: Job? get() = closeJob.load()
 
     /**
      * Round-5 concurrency #5: close() guard. A second concurrent invocation
      * (e.g. session close + read-loop death close racing) used to launch a
      * parallel teardown that called scope.cancel() and socket.close() while
-     * the first close was mid-joinAll. We now memoize the teardown Job so
-     * the second caller awaits the first's completion instead.
+     * the first close was mid-joinAll. We memoize the teardown Job so the
+     * second caller awaits the first's completion instead.
+     *
+     * Lock-free CAS replaces the previous `synchronized(this)` double-checked
+     * init: the close path is single-shot ("first writer wins"), which is
+     * exactly what `compareAndSet` expresses.
      */
-    @Volatile
-    private var closeJob: Job? = null
+    private val closeJob: AtomicReference<Job?> = AtomicReference(null)
 
     fun start() {
         connection.start()
@@ -365,38 +371,45 @@ class QuicConnectionDriver(
         // second concurrent caller (which is common: session.close() and
         // read-loop death both race to close()) awaits the same Job rather
         // than launching a parallel teardown.
-        if (closeJob != null) return
-        synchronized(this) {
-            if (closeJob != null) return
-            closeJob =
-                parentScope.launch {
-                    connection.close(0L, "")
-                    wakeup()
-                    val send = sendJob
-                    // Bounded wait for the send loop to flush CONNECTION_CLOSE.
-                    // We don't want to hang forever if the writer is wedged —
-                    // the timeout is the upper bound on how long close() blocks.
-                    withTimeoutOrNull(CLOSE_FLUSH_TIMEOUT_MILLIS) {
-                        // Spin until the writer has actually drained the queued
-                        // close. The CLOSING-status check transitions to CLOSED
-                        // once drainOutbound builds the CONNECTION_CLOSE packet.
-                        while (connection.status == QuicConnection.Status.CLOSING) {
-                            kotlinx.coroutines.delay(1)
-                        }
+        if (closeJob.load() != null) return
+        // Build the teardown coroutine LAZY so we can race-test the CAS
+        // without paying for a launched-and-cancelled Job on the loser.
+        val teardown =
+            parentScope.launch(start = kotlinx.coroutines.CoroutineStart.LAZY) {
+                connection.close(0L, "")
+                wakeup()
+                val send = sendJob
+                // Bounded wait for the send loop to flush CONNECTION_CLOSE.
+                // We don't want to hang forever if the writer is wedged —
+                // the timeout is the upper bound on how long close() blocks.
+                withTimeoutOrNull(CLOSE_FLUSH_TIMEOUT_MILLIS) {
+                    // Spin until the writer has actually drained the queued
+                    // close. The CLOSING-status check transitions to CLOSED
+                    // once drainOutbound builds the CONNECTION_CLOSE packet.
+                    while (connection.status == QuicConnection.Status.CLOSING) {
+                        kotlinx.coroutines.delay(1)
                     }
-                    // Now flip to CLOSED so both loops exit their while-guards.
-                    connection.markClosedExternally("driver close requested")
-                    wakeup()
-                    // Wait for both loops to actually exit — joinAll won't
-                    // return until the in-flight socket.send() completes.
-                    withTimeoutOrNull(CLOSE_FLUSH_TIMEOUT_MILLIS) {
-                        listOfNotNull(readJob, send).joinAll()
-                    }
-                    // Final teardown — cancel guarantees both jobs are done
-                    // before we close the socket.
-                    scope.cancel()
-                    socket.close()
                 }
+                // Now flip to CLOSED so both loops exit their while-guards.
+                connection.markClosedExternally("driver close requested")
+                wakeup()
+                // Wait for both loops to actually exit — joinAll won't
+                // return until the in-flight socket.send() completes.
+                withTimeoutOrNull(CLOSE_FLUSH_TIMEOUT_MILLIS) {
+                    listOfNotNull(readJob, send).joinAll()
+                }
+                // Final teardown — cancel guarantees both jobs are done
+                // before we close the socket.
+                scope.cancel()
+                socket.close()
+            }
+        if (closeJob.compareAndSet(null, teardown)) {
+            teardown.start()
+        } else {
+            // A concurrent close() already installed the teardown Job; drop
+            // ours without ever starting it. The winner's Job runs, this
+            // call is a no-op (matching the original idempotent contract).
+            teardown.cancel()
         }
     }
 

quic/src/commonMain/kotlin/com/vitorpamplona/quic/stream/SendBuffer.kt

@@ -60,9 +60,14 @@ package com.vitorpamplona.quic.stream
  * send loop under the connection mutex; [markAcked] / [markLost] run on
  * the parser path also under the connection mutex. The two execution
  * paths are NOT serialised by a shared lock, so all internal state is
- * mutated under `synchronized(this)`. Even the cheap getters
- * ([readableBytes], [sentOffset], [finPending], [finSent]) take the
- * monitor so a writer pre-flight check can't observe torn state.
+ * mutated under `synchronized(this)`. Single-field reads
+ * ([nextOffset], [sentOffset], [finPending], [finSent], [finAcked])
+ * use `@Volatile` backing fields and bypass the monitor — they cannot
+ * tear (Boolean is single-byte; Long writes happen inside the
+ * synchronized block on JVM/Android, where `@Volatile Long` is
+ * atomic). [readableBytes] still synchronizes because its formula
+ * combines two fields and would otherwise observe a transient
+ * negative value if read mid-`takeChunk`.
  *
  * # FIN
  *
@@ -105,6 +110,7 @@ class SendBuffer(
     private var flushedFloor: Long = 0L
 
     /** Logical offset just past the last byte. Advances on [enqueue]. */
+    @Volatile
     private var _nextOffset: Long = 0L
 
     /**
@@ -115,6 +121,7 @@ class SendBuffer(
      *
      * Invariant: `flushedFloor <= nextSendOffset <= nextOffset`.
      */
+    @Volatile
     private var nextSendOffset: Long = 0L
 
     /**
@@ -151,14 +158,19 @@ class SendBuffer(
      */
     private var retransmitTotalBytes: Long = 0L
 
+    @Volatile
     private var _finPending: Boolean = false
+
+    @Volatile
     private var _finSent: Boolean = false
+
+    @Volatile
     private var _finAcked: Boolean = false
 
-    val nextOffset: Long get() = synchronized(this) { _nextOffset }
-    val finPending: Boolean get() = synchronized(this) { _finPending }
-    val finSent: Boolean get() = synchronized(this) { _finSent }
-    val finAcked: Boolean get() = synchronized(this) { _finAcked }
+    val nextOffset: Long get() = _nextOffset
+    val finPending: Boolean get() = _finPending
+    val finSent: Boolean get() = _finSent
+    val finAcked: Boolean get() = _finAcked
 
     /**
      * Bytes the writer would emit on the next [takeChunk] before any
@@ -183,7 +195,7 @@ class SendBuffer(
      * the high-water mark of *fresh* sends only, not the cumulative
      * retransmit volume.
      */
-    val sentOffset: Long get() = synchronized(this) { nextSendOffset }
+    val sentOffset: Long get() = nextSendOffset
 
     fun enqueue(bytes: ByteArray) {
         if (bytes.isEmpty()) return

quic/src/jvmAndroid/kotlin/com/vitorpamplona/quic/crypto/PlatformCrypto.kt

@@ -26,11 +26,23 @@ import javax.crypto.spec.SecretKeySpec
 
 /**
  * One-block AES-ECB encryption via JCA. Used only by QUIC header protection
- * (one block per packet, so no need for a more elaborate API).
+ * (one block per packet). The `Cipher` instance is cached per-thread so the
+ * read/send loops avoid `Cipher.getInstance(...)` provider lookup on every
+ * packet — which is the dominant cost for a 16-byte one-shot AEAD-less call.
+ *
+ * `ThreadLocal` is safe here because the call is stateless: every invocation
+ * re-`init`s with the caller-supplied key before `doFinal`, so a coroutine
+ * that hops dispatchers between calls just lands on whichever thread's cached
+ * `Cipher` it ends up on. No state leaks across calls.
  */
+private val aesEcbCipher: ThreadLocal<Cipher> =
+    ThreadLocal.withInitial { Cipher.getInstance("AES/ECB/NoPadding") }
+
 actual val PlatformAesOneBlock: AesOneBlockEncrypt =
     AesOneBlockEncrypt { key, block ->
-        val cipher = Cipher.getInstance("AES/ECB/NoPadding")
+        // .get() is non-null because withInitial supplies a Cipher, but
+        // Kotlin sees the Java return type as platform-nullable.
+        val cipher = aesEcbCipher.get()!!
         cipher.init(Cipher.ENCRYPT_MODE, SecretKeySpec(key, "AES"))
         cipher.doFinal(block)
     }

quic/src/jvmAndroid/kotlin/com/vitorpamplona/quic/tls/JdkCertificateValidator.kt

@@ -191,10 +191,20 @@ class JdkCertificateValidator(
             // GeneralName type 2 = dNSName, type 7 = iPAddress.
             if (type == 2 && dnsMatches(idnAscii(value), normalizedHost)) return true
             if (type == 7 && hostAsIp != null) {
+                // Defense-in-depth: a malformed cert could put a hostname in
+                // a type 7 SAN. Without this gate, InetAddress.getByName(value)
+                // would perform a DNS A/AAAA lookup on the validation path,
+                // both leaking the name in plaintext and blocking the read
+                // loop on the system resolver. Forcing a literal check keeps
+                // the JDK call to pure parsing (no I/O, no name service).
                 val sanIp =
-                    try {
-                        InetAddress.getByName(value).hostAddress
-                    } catch (_: Throwable) {
+                    if (looksLikeIpLiteral(value)) {
+                        try {
+                            InetAddress.getByName(value).hostAddress
+                        } catch (_: Throwable) {
+                            null
+                        }
+                    } else {
                         null
                     }
                 if (sanIp != null && sanIp.equals(hostAsIp, ignoreCase = true)) return true