Serialize process-wide memory barriers under Rosetta

src/libraries/System.Threading/tests/InterlockedTests.cs

@@ -7,6 +7,7 @@
 using System.Runtime.CompilerServices;
 using System.Text;
 using System.Threading.Tasks;
+using Microsoft.DotNet.RemoteExecutor;
 using Xunit;
 
 namespace System.Threading.Tests
@@ -1492,6 +1493,41 @@ public void MemoryBarrierProcessWide()
             Assert.Equal(1000*1000, count);
         }
 
+        private static bool IsThreadingAndRemoteExecutorSupported =>
+            PlatformDetection.IsMultithreadingSupported && RemoteExecutor.IsSupported;
+
+        [ConditionalFact(typeof(InterlockedTests), nameof(IsThreadingAndRemoteExecutorSupported))]
+        public void MemoryBarrierProcessWide_ConcurrentCalls()
+        {
+            RemoteExecutor.Invoke(static () =>
+            {
+                const int ThreadCount = 4;
+                const int IterationCount = 100;
+
+                using Barrier startBarrier = new Barrier(ThreadCount);
+                Thread[] threads = new Thread[ThreadCount];
+
+                for (int i = 0; i < threads.Length; i++)
+                {
+                    threads[i] = new Thread(() =>
+                    {
+                        startBarrier.SignalAndWait();
+
+                        for (int j = 0; j < IterationCount; j++)
+                        {
+                            Interlocked.MemoryBarrierProcessWide();
+                        }
+                    });
+                    threads[i].Start();
+                }
+
+                foreach (Thread thread in threads)
+                {
+                    thread.Join();
+                }
+            }).Dispose();
+        }
+
         // Taking this lock on the same thread repeatedly is very fast because it has no interlocked operations.
         // Switching the thread where the lock is taken is expensive because of allocation and FlushProcessWriteBuffers.
         private class AsymmetricLock

src/native/minipal/memorybarrierprocesswide.c

@@ -17,6 +17,7 @@
 #ifdef __APPLE__
 #include <stdlib.h>
 #include <mach/thread_state.h>
+#include <minipal/cpufeatures.h>
 
 #define CHECK_MACH(_msg, machret) do {                                      \
         if (machret != KERN_SUCCESS)                                        \
@@ -40,6 +41,9 @@
 #include <sys/membarrier.h>
 #endif
 
+// Mutex to serialize implementations that require it
+static pthread_mutex_t g_flushProcessWriteBuffersMutex;
+
 #if HAVE_SYS_MEMBARRIER_H
 static bool CanFlushUsingMembarrier(void)
 {
@@ -79,10 +83,9 @@ static bool s_flushUsingMemBarrier = false;
 // Helper memory page used by the fallback path
 static uint8_t* g_helperPage = NULL;
 
-// Mutex to make the fallback path thread safe
-static pthread_mutex_t g_flushProcessWriteBuffersMutex;
-
 static size_t s_pageSize = 0;
+#else
+static bool s_serializeAppleMemoryBarrier = false;
 #endif // !HOST_APPLE
 #endif // !HOST_WASM
 
@@ -99,6 +102,16 @@ bool minipal_initialize_memory_barrier_process_wide(void)
     // browser/wasm is currently single threaded
 #elif defined(HOST_APPLE)
     // Apple platforms do not support membarrier, so we use a different mechanism
+    if (minipal_detect_rosetta())
+    {
+        int status = pthread_mutex_init(&g_flushProcessWriteBuffersMutex, NULL);
+        if (status != 0)
+        {
+            return false;
+        }
+
+        s_serializeAppleMemoryBarrier = true;
+    }
 #else
 #if HAVE_SYS_MEMBARRIER_H
     if (CanFlushUsingMembarrier())
@@ -155,6 +168,14 @@ void minipal_memory_barrier_process_wide(void)
 #ifdef HOST_WASM
     // browser/wasm is currently single threaded
 #elif defined(HOST_APPLE)
+    int status;
+    if (s_serializeAppleMemoryBarrier)
+    {
+        status = pthread_mutex_lock(&g_flushProcessWriteBuffersMutex);
+        (void)status; // unused in release config
+        assert(status == 0 && "Failed to lock the flushProcessWriteBuffersMutex lock");
+    }
+
     mach_msg_type_number_t cThreads;
     thread_act_t *pThreads;
     kern_return_t machret = task_threads(mach_task_self(), &pThreads, &cThreads);
@@ -181,6 +202,13 @@ void minipal_memory_barrier_process_wide(void)
     // Deallocate the thread list now we're done with it.
     machret = vm_deallocate(mach_task_self(), (vm_address_t)pThreads, cThreads * sizeof(thread_act_t));
     CHECK_MACH("vm_deallocate()", machret);
+
+    if (s_serializeAppleMemoryBarrier)
+    {
+        status = pthread_mutex_unlock(&g_flushProcessWriteBuffersMutex);
+        (void)status; // unused in release config
+        assert(status == 0 && "Failed to unlock the flushProcessWriteBuffersMutex lock");
+    }
 #else // !HOST_APPLE && !HOST_WASM
 #if HAVE_SYS_MEMBARRIER_H
     if (s_flushUsingMemBarrier)