Unix.fork: do not destroy IO channel mutexes in the child process

Some of these mutexes may be locked, so this is undefined behavior,
and can cause fatal errors in the C thread library (ocaml#12636).

Leaving IO channel mutexes unchanged is what OCaml 5 does.

otherlibs/systhreads/st_stubs.c

@@ -401,8 +401,6 @@ static void caml_thread_remove_info(caml_thread_t th)
 
 static void caml_thread_reinitialize(void)
 {
-  struct channel * chan;
-
   /* Remove all other threads (now nonexistent)
      from the doubly-linked list of threads */
   while (curr_thread->next != curr_thread) {
@@ -417,15 +415,13 @@ static void caml_thread_reinitialize(void)
   /* Tick thread is not currently running in child process, will be
      re-created at next Thread.create */
   caml_tick_thread_running = 0;
-  /* Destroy all IO mutexes; will be reinitialized on demand */
-  for (chan = caml_all_opened_channels;
-       chan != NULL;
-       chan = chan->next) {
-    if (chan->mutex != NULL) {
-      st_mutex_destroy(chan->mutex);
-      chan->mutex = NULL;
-    }
-  }
+  /* IO channel mutexes that are locked at the time of the fork()
+     remain locked forever in the child process.  This can cause
+     deadlocks.  We used to destroy all IO channel mutexes, so that
+     they are reinitialized on demand and deadlocks are avoided.
+     However, this could cause fatal errors in the C thread library
+     (see #12636).  So, like in OCaml 5, let's keep the IO channel
+     mutexes unchanged. */
 }
 
 /* Initialize the thread machinery */