diff --git a/src/egl/main/eglapi.c b/src/egl/main/eglapi.c
index d308f08cdf6..de2c53eafd9 100644
--- a/src/egl/main/eglapi.c
+++ b/src/egl/main/eglapi.c
@@ -356,6 +356,7 @@ eglGetDisplay(EGLNativeDisplayType nativeDisplay)
    _EGLDisplay *disp;
    void *native_display_ptr;
 
+   util_perfetto_init();
    _EGL_FUNC_START(NULL, EGL_OBJECT_THREAD_KHR, NULL);
 
    STATIC_ASSERT(sizeof(void*) == sizeof(nativeDisplay));
@@ -420,6 +421,7 @@ eglGetPlatformDisplayEXT(EGLenum platform, void *native_display,
    EGLAttrib *attrib_list;
    EGLDisplay disp;
 
+   util_perfetto_init();
    _EGL_FUNC_START(NULL, EGL_OBJECT_THREAD_KHR, NULL);
 
    if (_eglConvertIntsToAttribs(int_attribs, &attrib_list) != EGL_SUCCESS)
@@ -434,6 +436,7 @@ EGLDisplay EGLAPIENTRY
 eglGetPlatformDisplay(EGLenum platform, void *native_display,
                       const EGLAttrib *attrib_list)
 {
+   util_perfetto_init();
    _EGL_FUNC_START(NULL, EGL_OBJECT_THREAD_KHR, NULL);
    return _eglGetPlatformDisplayCommon(platform, native_display, attrib_list);
 }
diff --git a/src/util/perf/cpu_trace.h b/src/util/perf/cpu_trace.h
index 75afb8cbd0a..c13a3821158 100644
--- a/src/util/perf/cpu_trace.h
+++ b/src/util/perf/cpu_trace.h
@@ -6,25 +6,46 @@
 #ifndef CPU_TRACE_H
 #define CPU_TRACE_H
 
+#include "u_perfetto.h"
+
+#include "util/macros.h"
+
+#if defined(HAVE_PERFETTO)
+
+/* note that util_perfetto_is_category_enabled always returns false util
+ * util_perfetto_init is called
+ */
+#define _MESA_TRACE_BEGIN(category, name)                                    \
+   do {                                                                      \
+      if (unlikely(util_perfetto_is_category_enabled(category)))             \
+         util_perfetto_trace_begin(category, name);                          \
+   } while (0)
+
+#define _MESA_TRACE_END(category)                                            \
+   do {                                                                      \
+      if (unlikely(util_perfetto_is_category_enabled(category)))             \
+         util_perfetto_trace_end(category);                                  \
+   } while (0)
+
 /* NOTE: for now disable atrace for C++ to workaround a ndk bug with ordering
  * between stdatomic.h and atomic.h.  See:
  *
  *   https://github.com/android/ndk/issues/1178
  */
-#if defined(ANDROID) && !defined(__cplusplus)
+#elif defined(ANDROID) && !defined(__cplusplus)
 
 #include <cutils/trace.h>
 
-#define MESA_TRACE_BEGIN(name) atrace_begin(ATRACE_TAG_GRAPHICS, name)
-#define MESA_TRACE_END() atrace_end(ATRACE_TAG_GRAPHICS)
+#define _MESA_TRACE_BEGIN(category, name)                                    \
+   atrace_begin(ATRACE_TAG_GRAPHICS, name)
+#define _MESA_TRACE_END(category) atrace_end(ATRACE_TAG_GRAPHICS)
 
 #else
 
-/* XXX we would like to use perfetto, but it lacks a C header */
-#define MESA_TRACE_BEGIN(name)
-#define MESA_TRACE_END()
+#define _MESA_TRACE_BEGIN(category, name)
+#define _MESA_TRACE_END(category)
 
-#endif /* ANDROID */
+#endif /* HAVE_PERFETTO */
 
 #if __has_attribute(cleanup) && __has_attribute(unused)
 
@@ -35,34 +56,54 @@
 /* This must expand to a single non-scoped statement for
  *
  *    if (cond)
- *       MESA_TRACE_SCOPE(...)
+ *       _MESA_TRACE_SCOPE(...)
  *
  * to work.
  */
-#define MESA_TRACE_SCOPE(name)                                               \
+#define _MESA_TRACE_SCOPE(category, name)                                    \
    int _MESA_TRACE_SCOPE_VAR(__LINE__)                                       \
-      __attribute__((cleanup(mesa_trace_scope_end), unused)) =               \
-         mesa_trace_scope_begin(name)
+      __attribute__((cleanup(_mesa_trace_scope_end), unused)) =              \
+         _mesa_trace_scope_begin(category, name)
 
 static inline int
-mesa_trace_scope_begin(const char *name)
+_mesa_trace_scope_begin(enum util_perfetto_category category,
+                        const char *name)
 {
-   MESA_TRACE_BEGIN(name);
-   return 0;
+   _MESA_TRACE_BEGIN(category, name);
+   return category;
 }
 
 static inline void
-mesa_trace_scope_end(int *scope)
+_mesa_trace_scope_end(int *scope)
 {
-   MESA_TRACE_END();
+   /* we save the category in the scope variable */
+   _MESA_TRACE_END(*scope);
 }
 
 #else
 
-#define MESA_TRACE_SCOPE(name)
+#define _MESA_TRACE_SCOPE(category, name)
 
 #endif /* __has_attribute(cleanup) && __has_attribute(unused) */
 
-#define MESA_TRACE_FUNC() MESA_TRACE_SCOPE(__func__)
+/* These use the default category.  Drivers or subsystems can use these, or
+ * define their own categories/macros.
+ */
+#define MESA_TRACE_BEGIN(name)                                               \
+   _MESA_TRACE_BEGIN(UTIL_PERFETTO_CATEGORY_DEFAULT, name)
+#define MESA_TRACE_END() _MESA_TRACE_END(UTIL_PERFETTO_CATEGORY_DEFAULT)
+#define MESA_TRACE_SCOPE(name)                                               \
+   _MESA_TRACE_SCOPE(UTIL_PERFETTO_CATEGORY_DEFAULT, name)
+#define MESA_TRACE_FUNC()                                                    \
+   _MESA_TRACE_SCOPE(UTIL_PERFETTO_CATEGORY_DEFAULT, __func__)
+
+/* these use the slow category */
+#define MESA_TRACE_BEGIN_SLOW(name)                                          \
+   _MESA_TRACE_BEGIN(UTIL_PERFETTO_CATEGORY_SLOW, name)
+#define MESA_TRACE_END_SLOW() _MESA_TRACE_END(UTIL_PERFETTO_CATEGORY_SLOW)
+#define MESA_TRACE_SCOPE_SLOW(name)                                          \
+   _MESA_TRACE_SCOPE(UTIL_PERFETTO_CATEGORY_SLOW, name)
+#define MESA_TRACE_FUNC_SLOW()                                               \
+   _MESA_TRACE_SCOPE(UTIL_PERFETTO_CATEGORY_SLOW, __func__)
 
 #endif /* CPU_TRACE_H */