Add config option to determine whether timerfd is used.

Add a new configuration option "reactor" / "use_timerfd" that is used by
the epoll_reactor. When true (the default), the reactor uses a timerfd
descriptor to manage timeouts. When false, the duration until the next
timeout is computed and passed as a timeout to epoll_wait.

Author: chriskohlhoff

include/asio/detail/impl/epoll_reactor.ipp

@@ -43,7 +43,8 @@ epoll_reactor::epoll_reactor(asio::execution_context& ctx)
         config(ctx).get("reactor", "registration_locking_spin_count", 0)),
     interrupter_(),
     epoll_fd_(do_epoll_create()),
-    timer_fd_(do_timerfd_create()),
+    timer_fd_(config(ctx).get("reactor", "use_timerfd", true)
+        ? do_timerfd_create() : -1),
     shutdown_(false),
     io_locking_(config(ctx).get("reactor", "io_locking", true)),
     io_locking_spin_count_(
@@ -109,9 +110,11 @@ void epoll_reactor::notify_fork(
     epoll_fd_ = do_epoll_create();
 
     if (timer_fd_ != -1)
+    {
       ::close(timer_fd_);
-    timer_fd_ = -1;
-    timer_fd_ = do_timerfd_create();
+      timer_fd_ = -1;
+      timer_fd_ = do_timerfd_create();
+    }
 
     interrupter_.recreate();
 

src/doc/overview/configuration.qbk

@@ -177,6 +177,19 @@ below.
       fails with `EAGAIN`.
     ]
   ]
+  [
+    [`reactor`]
+    [`use_timerfd`]
+    [`bool`]
+    [`true`]
+    [
+      Linux [^epoll] backend only.
+
+      When `true`, the reactor uses a [^timerfd] descriptor to manage timeouts.
+      When `false`, the duration until the next timeout is computed and passed
+      as a timeout to [^epoll_wait].
+    ]
+  ]
   [
     [`timer`]
     [`heap_reserve`]