diff --git a/compiler-rt/lib/scudo/scudo_tsd_shared.cpp b/compiler-rt/lib/scudo/scudo_tsd_shared.cpp
--- a/compiler-rt/lib/scudo/scudo_tsd_shared.cpp
+++ b/compiler-rt/lib/scudo/scudo_tsd_shared.cpp
@@ -17,6 +17,7 @@
 namespace __scudo {
 
 static pthread_once_t GlobalInitialized = PTHREAD_ONCE_INIT;
+atomic_uint32_t ScudoInitted;
 pthread_key_t PThreadKey;
 
 static atomic_uint32_t CurrentIndex;
@@ -32,7 +33,6 @@
 
 static void initOnce() {
   CHECK_EQ(pthread_key_create(&PThreadKey, NULL), 0);
-  initScudo();
   NumberOfTSDs = Min(Max(1U, GetNumberOfCPUsCached()),
                      static_cast<u32>(SCUDO_SHARED_TSD_POOL_SIZE));
   TSDs = reinterpret_cast<ScudoTSD *>(
@@ -57,6 +57,11 @@
 #endif  // SANITIZER_ANDROID
 }
 
+void InitScudoOnce() {
+  u32 Initted = atomic_fetch_add(&ScudoInitted, 1, memory_order_relaxed);
+  if (UNLIKELY(!Initted))
+    initScudo();
+}
 void initThread(bool MinimalInit) {
   pthread_once(&GlobalInitialized, initOnce);
   // Initial context assignment is done in a plain round-robin fashion.
diff --git a/compiler-rt/lib/scudo/scudo_tsd_shared.inc b/compiler-rt/lib/scudo/scudo_tsd_shared.inc
--- a/compiler-rt/lib/scudo/scudo_tsd_shared.inc
+++ b/compiler-rt/lib/scudo/scudo_tsd_shared.inc
@@ -17,6 +17,7 @@
 #if !SCUDO_TSD_EXCLUSIVE
 
 extern pthread_key_t PThreadKey;
+extern void InitScudoOnce();
 
 #if SANITIZER_LINUX && !SANITIZER_ANDROID
 __attribute__((tls_model("initial-exec")))
@@ -34,8 +35,9 @@
 }
 
 ALWAYS_INLINE void initThreadMaybe(bool MinimalInit = false) {
+  InitScudoOnce();
   if (LIKELY(getCurrentTSD()))
-    return;
+   return;
   initThread(MinimalInit);
 }