diff --git a/libc/stdio/local.h b/libc/stdio/local.h
index d306a21..1ecf122 100644
--- a/libc/stdio/local.h
+++ b/libc/stdio/local.h
@@ -204,6 +204,7 @@
 __LIBC32_LEGACY_PUBLIC__ int _fwalk(int (*)(FILE*));
 
 off64_t __sseek64(void*, off64_t, int);
+int __sflush_locked(FILE*);
 int __swhatbuf(FILE*, size_t*, int*);
 wint_t __fgetwc_unlock(FILE*);
 wint_t __ungetwc(wint_t, FILE*);
diff --git a/libc/stdio/refill.c b/libc/stdio/refill.c
index cfa2bfd..1df4191 100644
--- a/libc/stdio/refill.c
+++ b/libc/stdio/refill.c
@@ -40,7 +40,7 @@
 lflush(FILE *fp)
 {
 	if ((fp->_flags & (__SLBF|__SWR)) == (__SLBF|__SWR))
-		return (__sflush(fp));	/* ignored... */
+		return (__sflush_locked(fp));	/* ignored... */
 	return (0);
 }
 
diff --git a/libc/stdio/stdio.cpp b/libc/stdio/stdio.cpp
index 4cec757..91c7689 100644
--- a/libc/stdio/stdio.cpp
+++ b/libc/stdio/stdio.cpp
@@ -106,7 +106,7 @@
 FILE* stdout = &__sF[1];
 FILE* stderr = &__sF[2];
 
-static pthread_mutex_t __stdio_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
+static pthread_mutex_t __stdio_mutex = PTHREAD_MUTEX_INITIALIZER;
 
 static uint64_t __get_file_tag(FILE* fp) {
   // Don't use a tag for the standard streams.
@@ -211,21 +211,23 @@
 }
 
 int _fwalk(int (*callback)(FILE*)) {
-  pthread_mutex_lock(&__stdio_mutex);
   int result = 0;
   for (glue* g = &__sglue; g != nullptr; g = g->next) {
     FILE* fp = g->iobs;
     for (int n = g->niobs; --n >= 0; ++fp) {
-      ScopedFileLock sfl(fp);
       if (fp->_flags != 0 && (fp->_flags & __SIGN) == 0) {
         result |= (*callback)(fp);
       }
     }
   }
-  pthread_mutex_unlock(&__stdio_mutex);
   return result;
 }
 
+extern "C" __LIBC_HIDDEN__ void __libc_stdio_cleanup(void) {
+  // Equivalent to fflush(nullptr), but without all the locking since we're shutting down anyway.
+  _fwalk(__sflush);
+}
+
 static FILE* __fopen(int fd, int flags) {
 #if !defined(__LP64__)
   if (fd > SHRT_MAX) {
@@ -520,6 +522,11 @@
   return 0;
 }
 
+int __sflush_locked(FILE* fp) {
+  ScopedFileLock sfl(fp);
+  return __sflush(fp);
+}
+
 int __sread(void* cookie, char* buf, int n) {
   FILE* fp = reinterpret_cast<FILE*>(cookie);
   return TEMP_FAILURE_RETRY(read(fp->_file, buf, n));
@@ -1061,7 +1068,7 @@
 }
 
 static int fflush_all() {
-  return _fwalk(__sflush);
+  return _fwalk(__sflush_locked);
 }
 
 int fflush(FILE* fp) {
diff --git a/libc/stdlib/atexit.c b/libc/stdlib/atexit.c
index 692e0c0..0efb118 100644
--- a/libc/stdlib/atexit.c
+++ b/libc/stdlib/atexit.c
@@ -188,7 +188,8 @@
 
 	/* If called via exit(), flush output of all open files. */
 	if (dso == NULL) {
-		fflush(NULL);
+		extern void __libc_stdio_cleanup(void);
+		__libc_stdio_cleanup();
 	}
 
   /* BEGIN android-changed: call __unregister_atfork if dso is not null */