pal init: Init membarrier() before first thread to improve start time

Refactor InitializeFlushProcessWriteBuffers(): Split membarrier()
initialization into a new InitializeMembarrier() helper function.

InitializeMembarrier() earlier before first thread is created to improve
process start time on Linux. More details can be found in issue 106722.

Fixes #106722

src/coreclr/pal/src/include/pal/process.h

@@ -199,6 +199,15 @@ Return
 --*/
 BOOL InitializeFlushProcessWriteBuffers();
 
+/*++
+Function:
+  InitializeMembarrier
+
+Abstract
+  This function initializes the Linux "expedited membarrier" syscall, if available
+--*/
+void InitializeMembarrier();
+
 #ifdef __cplusplus
 }
 #endif // __cplusplus

src/coreclr/pal/src/init/pal.cpp

@@ -336,6 +336,10 @@ Initialize(
     }
 #endif // __APPLE__
 
+    // Init "expedited membarrier" as soon as possible (before worker
+    // threads are created) for faster process startup on Linux.
+    InitializeMembarrier();
+
     CriticalSectionSubSysInitialize();
 
     if(nullptr == init_critsec)

src/coreclr/pal/src/thread/process.cpp

@@ -167,9 +167,10 @@ enum membarrier_cmd
 };
 
 //
-// Tracks if the OS supports FlushProcessWriteBuffers using membarrier
+// Tracks if membarrier is initialized and available on the current system
 //
-static int s_flushUsingMemBarrier = 0;
+static BOOL s_membarrierInitialized = FALSE;
+static BOOL s_membarrierAvailable = FALSE;
 
 //
 // Helper memory page used by the FlushProcessWriteBuffers
@@ -2562,6 +2563,35 @@ PROCAbort(int signal, siginfo_t* siginfo)
     abort();
 }
 
+/*++
+Function:
+  InitializeMembarrier
+
+Abstract
+  This function initializes the Linux "expedited membarrier" syscall, if available
+--*/
+void
+InitializeMembarrier()
+{
+    if (!s_membarrierInitialized)
+    {
+        // Starting with Linux kernel 4.14, process memory barriers can be generated
+        // using MEMBARRIER_CMD_PRIVATE_EXPEDITED.
+        int mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
+        if (mask >= 0 &&
+            mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED)
+        {
+            // Register intent to use the private expedited command.
+            if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0) == 0)
+            {
+                s_membarrierAvailable = TRUE;
+            }
+        }
+    }
+
+    s_membarrierInitialized = TRUE;
+}
+
 /*++
 Function:
   InitializeFlushProcessWriteBuffers
@@ -2575,20 +2605,11 @@ BOOL
 InitializeFlushProcessWriteBuffers()
 {
     _ASSERTE(s_helperPage == 0);
-    _ASSERTE(s_flushUsingMemBarrier == 0);
 
-    // Starting with Linux kernel 4.14, process memory barriers can be generated
-    // using MEMBARRIER_CMD_PRIVATE_EXPEDITED.
-    int mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
-    if (mask >= 0 &&
-        mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED)
+    InitializeMembarrier();
+    if (s_membarrierAvailable)
     {
-        // Register intent to use the private expedited command.
-        if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0) == 0)
-        {
-            s_flushUsingMemBarrier = TRUE;
-            return TRUE;
-        }
+        return TRUE;
     }
 
 #ifdef TARGET_OSX
@@ -2645,7 +2666,7 @@ VOID
 PALAPI
 FlushProcessWriteBuffers()
 {
-    if (s_flushUsingMemBarrier)
+    if (s_membarrierAvailable)
     {
         int status = membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0);
         FATAL_ASSERT(status == 0, "Failed to flush using membarrier");