pal init: InitializeFlushProcessWriteBuffers() before first thread to…

… improve start time (#106724)

InitializeFlushProcessWriteBuffers() initializes expedited membarrier()
syscall on Linux, which is much slower when called in a multi-thread
process. Move this init earlier to improve dotnet process start time.
A detailed explanation can be found in issue 106722.

Fixes #106722

Co-authored-by: Haris Okanovic <harisokn@amazon.com>

src/coreclr/pal/inc/pal.h

@@ -191,6 +191,7 @@ PAL_IsDebuggerPresent();
 #define PAL_INITIALIZE_ENSURE_STACK_SIZE            0x20
 #define PAL_INITIALIZE_REGISTER_SIGNALS             0x40
 #define PAL_INITIALIZE_REGISTER_ACTIVATION_SIGNAL   0x80
+#define PAL_INITIALIZE_FLUSH_PROCESS_WRITE_BUFFERS  0x100
 
 // PAL_Initialize() flags
 #define PAL_INITIALIZE                 (PAL_INITIALIZE_SYNC_THREAD | \
@@ -206,7 +207,8 @@ PAL_IsDebuggerPresent();
                                         PAL_INITIALIZE_DEBUGGER_EXCEPTIONS | \
                                         PAL_INITIALIZE_ENSURE_STACK_SIZE | \
                                         PAL_INITIALIZE_REGISTER_SIGNALS | \
-                                        PAL_INITIALIZE_REGISTER_ACTIVATION_SIGNAL)
+                                        PAL_INITIALIZE_REGISTER_ACTIVATION_SIGNAL  | \
+                                        PAL_INITIALIZE_FLUSH_PROCESS_WRITE_BUFFERS)
 
 typedef DWORD (PALAPI_NOEXPORT *PTHREAD_START_ROUTINE)(LPVOID lpThreadParameter);
 typedef PTHREAD_START_ROUTINE LPTHREAD_START_ROUTINE;

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

@@ -382,6 +382,16 @@ Initialize(
             goto CLEANUP0a;
         }
 
+        if (flags & PAL_INITIALIZE_FLUSH_PROCESS_WRITE_BUFFERS)
+        {
+            // Initialize before first thread is created for faster load on Linux
+            if (!InitializeFlushProcessWriteBuffers())
+            {
+                palError = ERROR_PALINIT_INITIALIZE_FLUSH_PROCESS_WRITE_BUFFERS;
+                goto CLEANUP0a;
+            }
+        }
+
         // The gSharedFilesPath is allocated dynamically so its destructor does not get
         // called unexpectedly during cleanup
         gSharedFilesPath = InternalNew<PathCharString>();
@@ -787,11 +797,6 @@ PAL_InitializeCoreCLR(const char *szExePath, BOOL runningInExe)
         return ERROR_PALINIT_PROCABORT_INITIALIZE;
     }
 
-    if (!InitializeFlushProcessWriteBuffers())
-    {
-        return ERROR_PALINIT_INITIALIZE_FLUSH_PROCESS_WRITE_BUFFERS;
-    }
-
     return ERROR_SUCCESS;
 }