Delay Start Worker Threads

scripts/run-gtest.ps1

@@ -137,7 +137,10 @@ param (
     [string]$OsRunner = "",
 
     [Parameter(Mandatory = $false)]
-    [switch]$UseQtip = $false
+    [switch]$UseQtip = $false,
+
+    [Parameter(Mandatory = $false)]
+    [Int32]$Timeout = 60000
 )
 
 Set-StrictMode -Version 'Latest'
@@ -385,7 +388,7 @@ function Start-TestCase([String]$Name) {
 
     # Build up the argument list.
     $ResultsPath = Join-Path $LocalLogDir "results.xml"
-    $Arguments = "--gtest_catch_exceptions=0 --gtest_filter=$($Name) --gtest_output=xml:$($ResultsPath) --timeout 60000"
+    $Arguments = "--gtest_catch_exceptions=0 --gtest_filter=$($Name) --gtest_output=xml:$($ResultsPath) --timeout $Timeout"
     if ($BreakOnFailure) {
         $Arguments += " --gtest_break_on_failure"
     }

scripts/test.ps1

@@ -180,7 +180,10 @@ param (
     [string]$OsRunner = "",
 
     [Parameter(Mandatory = $false)]
-    [int]$NumIterations = 1
+    [int]$NumIterations = 1,
+
+    [Parameter(Mandatory = $false)]
+    [Int32]$Timeout = 60000
 )
 
 Set-StrictMode -Version 'Latest'
@@ -285,7 +288,7 @@ if (!(Test-Path $PfxFile)) {
 }
 
 # Build up all the arguments to pass to the Powershell script.
-$TestArguments =  "-IsolationMode $IsolationMode -PfxPath $PfxFile"
+$TestArguments =  "-IsolationMode $IsolationMode -PfxPath $PfxFile -Timeout $Timeout"
 
 if ($DuoNic) {
     $TestArguments += " -DuoNic"

src/core/worker.c

@@ -54,6 +54,42 @@
     _In_ QUIC_WORKER* Worker
     );
 
+_IRQL_requires_max_(PASSIVE_LEVEL)
+QUIC_STATUS
+QuicWorkerStartThread(
+    _In_ QUIC_WORKER* Worker
+    )
+{
+    const uint16_t ThreadFlags =
+        Worker->ExecProfile == QUIC_EXECUTION_PROFILE_TYPE_REAL_TIME ?
+            CXPLAT_THREAD_FLAG_SET_AFFINITIZE : CXPLAT_THREAD_FLAG_NONE;
+
+    CXPLAT_THREAD_CONFIG ThreadConfig = {
+        ThreadFlags,
+        QuicLibraryGetPartitionProcessor(Worker->PartitionIndex),
+        "quic_worker",
+        QuicWorkerThread,
+        Worker
+    };
+
+    QuicTraceLogVerbose(
+        WorkerStarting,
+        "[wrkr][%p] Starting thread",
+        Worker);
+
+    QUIC_STATUS Status = CxPlatThreadCreate(&ThreadConfig, &Worker->Thread);
+    if (QUIC_FAILED(Status)) {
+        QuicTraceEvent(
+            WorkerErrorStatus,
+            "[wrkr][%p] ERROR, %u, %s.",
+            Worker,
+            Status,
+            "CxPlatThreadCreate");
+    }
+
+    return Status;
+}
+
 _IRQL_requires_max_(PASSIVE_LEVEL)
 QUIC_STATUS
 QuicWorkerInitialize(
@@ -71,6 +107,7 @@
         Registration);
 
     Worker->Enabled = TRUE;
+    Worker->ExecProfile = ExecProfile;
     Worker->PartitionIndex = PartitionIndex;
     CxPlatDispatchLockInitialize(&Worker->Lock);
     CxPlatEventInitialize(&Worker->Done, TRUE, FALSE);
@@ -95,36 +132,24 @@
     Worker->ExecutionContext.NextTimeUs = UINT64_MAX;
     Worker->ExecutionContext.Ready = TRUE;
 
-#ifndef _KERNEL_MODE // Not supported on kernel mode
+#ifdef _KERNEL_MODE
+    //
+    // Kernel mode always has to start threads up front, because delay load
+    // requires creating the thread later at PASSIVE_LEVEL, but several of the
+    // places we would need to do so are at DISPATCH_LEVEL.
+    //
+    Status = QuicWorkerStartThread(Worker);
+#else
     if (ExecProfile != QUIC_EXECUTION_PROFILE_TYPE_MAX_THROUGHPUT) {
         Worker->IsExternal = TRUE;
-        CxPlatAddExecutionContext(&Worker->ExecutionContext, PartitionIndex);
-    } else
-#endif // _KERNEL_MODE
-    {
-        const uint16_t ThreadFlags =
-            ExecProfile == QUIC_EXECUTION_PROFILE_TYPE_REAL_TIME ?
-                CXPLAT_THREAD_FLAG_SET_AFFINITIZE : CXPLAT_THREAD_FLAG_NONE;
-
-        CXPLAT_THREAD_CONFIG ThreadConfig = {
-            ThreadFlags,
-            QuicLibraryGetPartitionProcessor(PartitionIndex),
-            "quic_worker",
-            QuicWorkerThread,
-            Worker
-        };
-
-        Status = CxPlatThreadCreate(&ThreadConfig, &Worker->Thread);
-        if (QUIC_FAILED(Status)) {
-            QuicTraceEvent(
-                WorkerErrorStatus,
-                "[wrkr][%p] ERROR, %u, %s.",
-                Worker,
-                Status,
-                "CxPlatThreadCreate");
-            goto Error;
-        }
+        CxPlatAddExecutionContext(&Worker->ExecutionContext, PartitionIndex, FALSE);
+    } else {
+        //
+        // Don't start the thread until it's needed.
+        //
+        Worker->DelayStart = TRUE;
     }
+#endif // _KERNEL_MODE
 
 Error:
 
@@ -225,6 +250,14 @@
 
     CxPlatDispatchLockAcquire(&Worker->Lock);
 
+#ifndef _KERNEL_MODE // Not supported on kernel mode
+    BOOLEAN DelayStart = FALSE;
+    if (Worker->DelayStart) {
+        DelayStart = TRUE;
+        Worker->DelayStart = FALSE;
+    }
+#endif // _KERNEL_MODE
+
     BOOLEAN WakeWorkerThread;
     if (!Connection->WorkerProcessing && !Connection->HasQueuedWork) {
         WakeWorkerThread = QuicWorkerIsIdle(Worker);
@@ -252,6 +285,12 @@
     if (WakeWorkerThread) {
         QuicWorkerThreadWake(Worker);
     }
+
+#ifndef _KERNEL_MODE
+    if (DelayStart) {
+        QuicWorkerStartThread(Worker);
+    }
+#endif // _KERNEL_MODE
 }
 
 _IRQL_requires_max_(DISPATCH_LEVEL)
@@ -265,6 +304,14 @@
 
     CxPlatDispatchLockAcquire(&Worker->Lock);
 
+#ifndef _KERNEL_MODE // Not supported on kernel mode
+    BOOLEAN DelayStart = FALSE;
+    if (Worker->DelayStart) {
+        DelayStart = TRUE;
+        Worker->DelayStart = FALSE;
+    }
+#endif // _KERNEL_MODE
+
     BOOLEAN WakeWorkerThread = QuicWorkerIsIdle(Worker);
 
     if (Connection->HasQueuedWork) {
@@ -283,6 +330,12 @@
     if (WakeWorkerThread) {
         QuicWorkerThreadWake(Worker);
     }
+
+#ifndef _KERNEL_MODE
+    if (DelayStart) {
+        QuicWorkerStartThread(Worker);
+    }
+#endif // _KERNEL_MODE
 }
 
 _IRQL_requires_max_(DISPATCH_LEVEL)
@@ -294,6 +347,14 @@
 {
     CxPlatDispatchLockAcquire(&Worker->Lock);
 
+#ifndef _KERNEL_MODE // Not supported on kernel mode
+    BOOLEAN DelayStart = FALSE;
+    if (Worker->DelayStart) {
+        DelayStart = TRUE;
+        Worker->DelayStart = FALSE;
+    }
+#endif // _KERNEL_MODE
+
     BOOLEAN WakeWorkerThread;
     if (Worker->OperationCount < MsQuicLib.Settings.MaxStatelessOperations &&
         QuicLibraryTryAddRefBinding(Operation->STATELESS.Context->Binding)) {
@@ -319,6 +380,12 @@
     } else if (WakeWorkerThread) {
         QuicWorkerThreadWake(Worker);
     }
+
+#ifndef _KERNEL_MODE
+    if (DelayStart) {
+        QuicWorkerStartThread(Worker);
+    }
+#endif // _KERNEL_MODE
 }
 
 _IRQL_requires_max_(PASSIVE_LEVEL)

src/core/worker.h

@@ -37,15 +37,25 @@ typedef struct QUIC_CACHEALIGN QUIC_WORKER {
     BOOLEAN IsActive;
 
     //
-    // The index into the partition array (of processors).
+    // TRUE if the worker still needs its thread to be started.
     //
-    uint16_t PartitionIndex;
+    BOOLEAN DelayStart;
+
+    //
+    // The execution profile for this worker, inherited from the registration.
+    //
+    QUIC_EXECUTION_PROFILE ExecProfile;
 
     //
     // The average queue delay connections experience, in microseconds.
     //
     uint32_t AverageQueueDelay;
 
+    //
+    // The index into the partition array (of processors).
+    //
+    uint16_t PartitionIndex;
+
     //
     // Timers for the worker's connections.
     //

src/generated/linux/platform_worker.c.clog.h

@@ -25,6 +25,24 @@
 #ifdef __cplusplus
 extern "C" {
 #endif
+/*----------------------------------------------------------
+// Decoder Ring for PlatformWorkerThreadStarting
+// [ lib][%p] Worker starting
+// QuicTraceLogInfo(
+        PlatformWorkerThreadStarting,
+        "[ lib][%p] Worker starting",
+        Worker);
+// arg2 = arg2 = Worker = arg2
+----------------------------------------------------------*/
+#ifndef _clog_3_ARGS_TRACE_PlatformWorkerThreadStarting
+#define _clog_3_ARGS_TRACE_PlatformWorkerThreadStarting(uniqueId, encoded_arg_string, arg2)\
+tracepoint(CLOG_PLATFORM_WORKER_C, PlatformWorkerThreadStarting , arg2);\
+
+#endif
+
+
+
+
 /*----------------------------------------------------------
 // Decoder Ring for PlatformWorkerThreadStart
 // [ lib][%p] Worker start
@@ -99,6 +117,26 @@ tracepoint(CLOG_PLATFORM_WORKER_C, LibraryError , arg2);\
 
 
 
+/*----------------------------------------------------------
+// Decoder Ring for LibraryErrorStatus
+// [ lib] ERROR, %u, %s.
+// QuicTraceEvent(
+            LibraryErrorStatus,
+            "[ lib] ERROR, %u, %s.",
+            Status,
+            "Create platform worker thread failed");
+// arg2 = arg2 = Status = arg2
+// arg3 = arg3 = "Create platform worker thread failed" = arg3
+----------------------------------------------------------*/
+#ifndef _clog_4_ARGS_TRACE_LibraryErrorStatus
+#define _clog_4_ARGS_TRACE_LibraryErrorStatus(uniqueId, encoded_arg_string, arg2, arg3)\
+tracepoint(CLOG_PLATFORM_WORKER_C, LibraryErrorStatus , arg2, arg3);\
+
+#endif
+
+
+
+
 #ifdef __cplusplus
 }
 #endif

src/generated/linux/platform_worker.c.clog.h.lttng.h

@@ -1,6 +1,25 @@
 
 
 
+/*----------------------------------------------------------
+// Decoder Ring for PlatformWorkerThreadStarting
+// [ lib][%p] Worker starting
+// QuicTraceLogInfo(
+        PlatformWorkerThreadStarting,
+        "[ lib][%p] Worker starting",
+        Worker);
+// arg2 = arg2 = Worker = arg2
+----------------------------------------------------------*/
+TRACEPOINT_EVENT(CLOG_PLATFORM_WORKER_C, PlatformWorkerThreadStarting,
+    TP_ARGS(
+        const void *, arg2), 
+    TP_FIELDS(
+        ctf_integer_hex(uint64_t, arg2, arg2)
+    )
+)
+
+
+
 /*----------------------------------------------------------
 // Decoder Ring for PlatformWorkerThreadStart
 // [ lib][%p] Worker start
@@ -78,3 +97,26 @@ TRACEPOINT_EVENT(CLOG_PLATFORM_WORKER_C, LibraryError,
         ctf_string(arg2, arg2)
     )
 )
+
+
+
+/*----------------------------------------------------------
+// Decoder Ring for LibraryErrorStatus
+// [ lib] ERROR, %u, %s.
+// QuicTraceEvent(
+            LibraryErrorStatus,
+            "[ lib] ERROR, %u, %s.",
+            Status,
+            "Create platform worker thread failed");
+// arg2 = arg2 = Status = arg2
+// arg3 = arg3 = "Create platform worker thread failed" = arg3
+----------------------------------------------------------*/
+TRACEPOINT_EVENT(CLOG_PLATFORM_WORKER_C, LibraryErrorStatus,
+    TP_ARGS(
+        unsigned int, arg2,
+        const char *, arg3), 
+    TP_FIELDS(
+        ctf_integer(unsigned int, arg2, arg2)
+        ctf_string(arg3, arg3)
+    )
+)