diff --git a/util/dprintf.c b/util/dprintf.c
index 34941f7..098988c 100644
--- a/util/dprintf.c
+++ b/util/dprintf.c
@@ -9,6 +9,8 @@
 
 #include "util/dprintf.h"
 
+static long dbg_buf_lock_init;
+static CRITICAL_SECTION dbg_buf_lock;
 static char dbg_buf[16384];
 static size_t dbg_buf_pos;
 
@@ -23,6 +25,26 @@ void dprintf(const char *fmt, ...)
 
 void dprintfv(const char *fmt, va_list ap)
 {
+    long init;
+
+    /* Static constructors in C are difficult to do in a way that works under
+       both GCC and MSVC, so we have to use atomic ops to ensure that the
+       buffer mutex is correctly initialized instead. */
+
+    do {
+        init = InterlockedCompareExchange(&dbg_buf_lock_init, 0, 1);
+
+        if (init == 0) {
+            /* We won the init race, global variable is now set to 1, other
+               threads will spin until it becomes -1. */
+            InitializeCriticalSection(&dbg_buf_lock);
+            dbg_buf_lock_init = -1;
+            init = -1;
+        }
+    } while (init >= 0);
+
+    EnterCriticalSection(&dbg_buf_lock);
+
     dbg_buf_pos += vsnprintf_s(
             dbg_buf + dbg_buf_pos,
             sizeof(dbg_buf) - dbg_buf_pos,
@@ -34,13 +56,13 @@ void dprintfv(const char *fmt, va_list ap)
         abort();
     }
 
-    if (strchr(dbg_buf, '\n') == NULL) {
-        return;
+    if (strchr(dbg_buf, '\n') != NULL) {
+        OutputDebugStringA(dbg_buf);
+        dbg_buf_pos = 0;
+        dbg_buf[0] = '\0';
     }
 
-    OutputDebugStringA(dbg_buf);
-    dbg_buf_pos = 0;
-    dbg_buf[0] = '\0';
+    LeaveCriticalSection(&dbg_buf_lock);
 }
 
 void dwprintf(const wchar_t *fmt, ...)