diff --git a/libs/binder/ProcessState.cpp b/libs/binder/ProcessState.cpp
index a42ede2..7c29dba 100644
--- a/libs/binder/ProcessState.cpp
+++ b/libs/binder/ProcessState.cpp
@@ -429,8 +429,17 @@
 }
 
 size_t ProcessState::getThreadPoolMaxTotalThreadCount() const {
+    // Need to read `mKernelStartedThreads` before `mThreadPoolStarted` (with
+    // non-relaxed memory ordering) to avoid a race like the following:
+    //
+    // thread A: if (mThreadPoolStarted) { // evaluates false
+    // thread B: mThreadPoolStarted = true;
+    // thread B: mKernelStartedThreads++;
+    // thread A: size_t kernelStarted = mKernelStartedThreads;
+    // thread A: LOG_ALWAYS_FATAL_IF(kernelStarted != 0, ...);
+    size_t kernelStarted = mKernelStartedThreads;
+
     if (mThreadPoolStarted) {
-        size_t kernelStarted = mKernelStartedThreads;
         size_t max = mMaxThreads;
         size_t current = mCurrentThreads;
 
@@ -460,7 +469,6 @@
 
     // must not be initialized or maybe has poll thread setup, we
     // currently don't track this in libbinder
-    size_t kernelStarted = mKernelStartedThreads;
     LOG_ALWAYS_FATAL_IF(kernelStarted != 0, "Expecting 0 kernel started threads but have %zu",
                         kernelStarted);
     return mCurrentThreads;
diff --git a/libs/binder/include/binder/ProcessState.h b/libs/binder/include/binder/ProcessState.h
index 021bd58..ffa2229 100644
--- a/libs/binder/include/binder/ProcessState.h
+++ b/libs/binder/include/binder/ProcessState.h
@@ -188,7 +188,7 @@
     Vector<handle_entry> mHandleToObject;
 
     bool mForked;
-    bool mThreadPoolStarted;
+    std::atomic_bool mThreadPoolStarted;
     volatile int32_t mThreadPoolSeq;
 
     CallRestriction mCallRestriction;