macOS: extend the workaround to cover the dyld/exc_server deadlock issue, since 12.1

Later, we should probably switch to using mach_exc_server generated from
`mig mach_exc.defs`.

(cherry picked from commit 267b124)

Author: vtjnash

src/signals-mach.c

@@ -526,27 +526,30 @@ static kern_return_t profiler_segv_handler
 }
 #endif
 
-static int jl_lock_profile_mach(void)
+// WARNING: we are unable to handle sigsegv while the dlsymlock is held
+static int jl_lock_profile_mach(int dlsymlock)
 {
     jl_lock_profile();
+    // workaround for old keymgr bugs
     void *unused = NULL;
     int keymgr_locked = _keymgr_get_and_lock_processwide_ptr_2(KEYMGR_GCC3_DW2_OBJ_LIST, &unused) == 0;
-    if (_dyld_atfork_prepare != NULL && _dyld_atfork_parent != NULL)
+    // workaround for new dlsym4 bugs (API and bugs introduced in macOS 12.1)
+    if (dlsymlock && _dyld_atfork_prepare != NULL && _dyld_atfork_parent != NULL)
         _dyld_atfork_prepare();
     return keymgr_locked;
 }
 
-static void jl_unlock_profile_mach(int keymgr_locked)
+static void jl_unlock_profile_mach(int dlsymlock, int keymgr_locked)
 {
-    if (_dyld_atfork_prepare != NULL && _dyld_atfork_parent != NULL)
-        _dyld_atfork_parent();
+    if (dlsymlock && _dyld_atfork_prepare != NULL && _dyld_atfork_parent != NULL) \
+        _dyld_atfork_parent(); \
     if (keymgr_locked)
         _keymgr_unlock_processwide_ptr(KEYMGR_GCC3_DW2_OBJ_LIST);
     jl_unlock_profile();
 }
 
-#define jl_lock_profile()       int keymgr_locked = jl_lock_profile_mach()
-#define jl_unlock_profile()     jl_unlock_profile_mach(keymgr_locked)
+#define jl_lock_profile()       int keymgr_locked = jl_lock_profile_mach(1)
+#define jl_unlock_profile()     jl_unlock_profile_mach(1, keymgr_locked)
 
 void *mach_profile_listener(void *arg)
 {
@@ -565,17 +568,21 @@ void *mach_profile_listener(void *arg)
         HANDLE_MACH_ERROR("mach_msg", ret);
         // sample each thread, round-robin style in reverse order
         // (so that thread zero gets notified last)
-        jl_lock_profile();
+        int keymgr_locked = jl_lock_profile_mach(0);
         for (i = jl_n_threads; i-- > 0; ) {
             // if there is no space left, break early
             if (jl_profile_is_buffer_full()) {
                 jl_profile_stop_timer();
                 break;
             }
 
+            if (_dyld_atfork_prepare != NULL && _dyld_atfork_parent != NULL)
+                _dyld_atfork_prepare(); // briefly acquire the dlsym lock
             host_thread_state_t state;
             jl_thread_suspend_and_get_state2(i, &state);
             unw_context_t *uc = (unw_context_t*)&state;
+            if (_dyld_atfork_prepare != NULL && _dyld_atfork_parent != NULL)
+                _dyld_atfork_parent(); // quickly release the dlsym lock
 
             if (running) {
 #ifdef LLVMLIBUNWIND
@@ -618,7 +625,7 @@ void *mach_profile_listener(void *arg)
             // We're done! Resume the thread.
             jl_thread_resume(i, 0);
         }
-        jl_unlock_profile();
+        jl_unlock_profile_mach(0, keymgr_locked);
         if (running) {
             // Reset the alarm
             kern_return_t ret = clock_alarm(clk, TIME_RELATIVE, timerprof, profile_port);