CDRIVER-4241 SHM counter on mac m1 arm64 implementation. (#894)

Co-authored-by: Kevin Albertson <[email protected]>

Author: devnexen

src/libmongoc/CMakeLists.txt

@@ -380,13 +380,16 @@ if (ENABLE_SHM_COUNTERS STREQUAL "AUTO")
    if (UNIX AND NOT APPLE)
       set (ENABLE_SHM_COUNTERS ON)
    endif ()
+   if (APPLE AND ${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm64")
+      set (ENABLE_SHM_COUNTERS ON)
+   endif ()
 endif ()
 
 if (ENABLE_SHM_COUNTERS STREQUAL "ON")
-   if (APPLE OR NOT UNIX)
+   if (( NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm64" AND APPLE ) OR NOT UNIX)
       message (
          FATAL_ERROR
-         "Shared memory performance counters not supported on Mac or Windows")
+	 "Shared memory performance counters only supported on arm64 Mac and Linux")
    endif ()
    set (MONGOC_ENABLE_SHM_COUNTERS 1)
    find_library(RT_LIBRARY rt HINTS /usr/lib32)

src/libmongoc/doc/basic-troubleshooting.rst

@@ -23,7 +23,7 @@ If available, the counters can be accessed outside of the application process vi
 This means that you can graph statistics about your application process easily from tools like Munin or Nagios.
 Your author often uses ``watch --interval=0.5 -d mongoc-stat $PID`` to monitor an application.
 
-Performance counters are only available on Linux platforms supporting shared memory segments.
+Performance counters are only available on Linux platforms and macOS arm64 platforms supporting shared memory segments.
 On supported platforms they are enabled by default.
 Applications can be built without the counters by specifying the cmake option ``-DENABLE_SHM_COUNTERS=OFF``. Additionally, if
 performance counters are already compiled, they can be disabled at runtime by specifying the environment variable ``MONGOC_DISABLE_SHM``.

src/libmongoc/src/mongoc/mongoc-counters-private.h

@@ -101,6 +101,19 @@ _mongoc_sched_getcpu (void)
 }
 #elif defined(MONGOC_HAVE_SCHED_GETCPU)
 #define _mongoc_sched_getcpu sched_getcpu
+#elif defined(__APPLE__) && defined(__aarch64__)
+static BSON_INLINE unsigned
+_mongoc_sched_getcpu (void)
+{
+   uintptr_t tls;
+   unsigned core_id;
+   /* Get the current thread ID, not the core ID.
+    * Getting the core ID requires privileged execution. */
+   __asm__ volatile("mrs %x0, tpidrro_el0" : "=r"(tls));
+   /* In ARM, only 8 cores are manageable. */
+   core_id = tls & 0x07u;
+   return core_id;
+}
 #else
 #define _mongoc_sched_getcpu() (0)
 #endif