*** empty log message ***

4 files changed, 99 insertions, 66 deletions

diff --git a/ev.c b/ev.c
index 6792e3f..b8dc960 100644
--- a/ev.c
+++ b/ev.c
@@ -1370,27 +1370,36 @@ evpipe_write (EV_P_ EV_ATOMIC_T *flag)
 {
   if (!*flag)
     {
-      int old_errno = errno; /* save errno because write might clobber it */
-      char dummy;
-
       *flag = 1;
 
-#if EV_USE_EVENTFD
-      if (evfd >= 0)
+      pipe_write_skipped = 1;
+
+      if (pipe_write_wanted)
         {
-          uint64_t counter = 1;
-          write (evfd, &counter, sizeof (uint64_t));
-        }
-      else
+          int old_errno = errno; /* save errno because write will clobber it */
+          char dummy;
+
+          pipe_write_skipped = 0;
+
+#if EV_USE_EVENTFD
+          if (evfd >= 0)
+            {
+              uint64_t counter = 1;
+              write (evfd, &counter, sizeof (uint64_t));
+            }
+          else
 #endif
-        /* win32 people keep sending patches that change this write() to send() */
-        /* and then run away. but send() is wrong, it wants a socket handle on win32 */
-        /* so when you think this write should be a send instead, please find out */
-        /* where your send() is from - it's definitely not the microsoft send, and */
-        /* tell me. thank you. */
-        write (evpipe [1], &dummy, 1);
+            {
+              /* win32 people keep sending patches that change this write() to send() */
+              /* and then run away. but send() is wrong, it wants a socket handle on win32 */
+              /* so when you think this write should be a send instead, please find out */
+              /* where your send() is from - it's definitely not the microsoft send, and */
+              /* tell me. thank you. */
+              write (evpipe [1], &dummy, 1);
+            }
 
-      errno = old_errno;
+          errno = old_errno;
+        }
     }
 }
 
@@ -1401,20 +1410,25 @@ pipecb (EV_P_ ev_io *iow, int revents)
 {
   int i;
 
-#if EV_USE_EVENTFD
-  if (evfd >= 0)
+  if (revents & EV_READ)
     {
-      uint64_t counter;
-      read (evfd, &counter, sizeof (uint64_t));
-    }
-  else
+#if EV_USE_EVENTFD
+      if (evfd >= 0)
+        {
+          uint64_t counter;
+          read (evfd, &counter, sizeof (uint64_t));
+        }
+      else
 #endif
-    {
-      char dummy;
-      /* see discussion in evpipe_write when you think this read should be recv in win32 */
-      read (evpipe [0], &dummy, 1);
+        {
+          char dummy;
+          /* see discussion in evpipe_write when you think this read should be recv in win32 */
+          read (evpipe [0], &dummy, 1);
+        }
     }
 
+  pipe_write_skipped = 0;
+
 #if EV_SIGNAL_ENABLE
   if (sig_pending)
     {
@@ -1453,6 +1467,8 @@ ev_feed_signal (int signum)
     return;
 #endif
 
+  evpipe_init (EV_A);
+
   signals [signum - 1].pending = 1;
   evpipe_write (EV_A_ &sig_pending);
 }
@@ -1759,27 +1775,29 @@ loop_init (EV_P_ unsigned int flags)
           && getenv ("LIBEV_FLAGS"))
         flags = atoi (getenv ("LIBEV_FLAGS"));
 
-      ev_rt_now         = ev_time ();
-      mn_now            = get_clock ();
-      now_floor         = mn_now;
-      rtmn_diff         = ev_rt_now - mn_now;
+      ev_rt_now          = ev_time ();
+      mn_now             = get_clock ();
+      now_floor          = mn_now;
+      rtmn_diff          = ev_rt_now - mn_now;
 #if EV_FEATURE_API
-      invoke_cb         = ev_invoke_pending;
+      invoke_cb          = ev_invoke_pending;
 #endif
 
-      io_blocktime      = 0.;
-      timeout_blocktime = 0.;
-      backend           = 0;
-      backend_fd        = -1;
-      sig_pending       = 0;
+      io_blocktime       = 0.;
+      timeout_blocktime  = 0.;
+      backend            = 0;
+      backend_fd         = -1;
+      sig_pending        = 0;
 #if EV_ASYNC_ENABLE
-      async_pending     = 0;
+      async_pending      = 0;
 #endif
+      pipe_write_skipped = 0;
+      pipe_write_wanted  = 0;
 #if EV_USE_INOTIFY
-      fs_fd             = flags & EVFLAG_NOINOTIFY ? -1 : -2;
+      fs_fd              = flags & EVFLAG_NOINOTIFY ? -1 : -2;
 #endif
 #if EV_USE_SIGNALFD
-      sigfd             = flags & EVFLAG_SIGNALFD  ? -2 : -1;
+      sigfd              = flags & EVFLAG_SIGNALFD  ? -2 : -1;
 #endif
 
       if (!(flags & EVBACKEND_MASK))
@@ -1954,12 +1972,7 @@ loop_fork (EV_P)
 
   if (ev_is_active (&pipe_w))
     {
-      /* this "locks" the handlers against writing to the pipe */
-      /* while we modify the fd vars */
-      sig_pending   = 1;
-#if EV_ASYNC_ENABLE
-      async_pending = 1;
-#endif
+      /* pipe_write_wanted must be false now, so modifying fd vars should be safe */
 
       ev_ref (EV_A);
       ev_io_stop (EV_A_ &pipe_w);
@@ -2501,7 +2514,10 @@ ev_run (EV_P_ int flags)
         /* update time to cancel out callback processing overhead */
         time_update (EV_A_ 1e100);
 
-        if (expect_true (!(flags & EVRUN_NOWAIT || idleall || !activecnt)))
+        /* from now on, we want a pipe-wake-up */
+        pipe_write_wanted = 1;
+
+        if (expect_true (!(flags & EVRUN_NOWAIT || idleall || !activecnt || pipe_write_skipped)))
           {
             waittime = MAX_BLOCKTIME;
 
@@ -2551,6 +2567,15 @@ ev_run (EV_P_ int flags)
         backend_poll (EV_A_ waittime);
         assert ((loop_done = EVBREAK_CANCEL, 1)); /* assert for side effect */
 
+        pipe_write_wanted = 0;
+
+        if (pipe_write_skipped)
+          {
+            assert (("libev: pipe_w not active, but pipe not written", ev_is_active (&pipe_w)));
+            ev_feed_event (EV_A_ &pipe_w, EV_CUSTOM);
+          }
+
+
         /* update ev_rt_now, do magic */
         time_update (EV_A_ waittime + sleeptime);
       }
diff --git a/ev.pod b/ev.pod
index 32e9479..ef5b4f6 100644
--- a/ev.pod
+++ b/ev.pod
@@ -610,11 +610,11 @@ hacks).
 
 On the negative side, the interface is I<bizarre> - so bizarre that
 even sun itself gets it wrong in their code examples: The event polling
-function sometimes returning events to the caller even though an error
+function sometimes returns events to the caller even though an error
 occurred, but with no indication whether it has done so or not (yes, it's
-even documented that way) - deadly for edge-triggered interfaces where
-you absolutely have to know whether an event occurred or not because you
-have to re-arm the watcher.
+even documented that way) - deadly for edge-triggered interfaces where you
+absolutely have to know whether an event occurred or not because you have
+to re-arm the watcher.
 
 Fortunately libev seems to be able to work around these idiocies.
 
@@ -2025,7 +2025,7 @@ do stuff) the timer will not fire more than once per event loop iteration.
 
 =item ev_timer_again (loop, ev_timer *)
 
-This will act as if the timer timed out and restart it again if it is
+This will act as if the timer timed out and restarts it again if it is
 repeating. The exact semantics are:
 
 If the timer is pending, its pending status is cleared.
@@ -3222,9 +3222,6 @@ of "global async watchers" by using a watcher on an otherwise unused
 signal, and C<ev_feed_signal> to signal this watcher from another thread,
 even without knowing which loop owns the signal.
 
-Unlike C<ev_signal> watchers, C<ev_async> works with any event loop, not
-just the default loop.
-
 =head3 Queueing
 
 C<ev_async> does not support queueing of data in any way. The reason
@@ -3333,13 +3330,16 @@ signal or similar contexts (see the discussion of C<EV_ATOMIC_T> in the
 embedding section below on what exactly this means).
 
 Note that, as with other watchers in libev, multiple events might get
-compressed into a single callback invocation (another way to look at this
-is that C<ev_async> watchers are level-triggered, set on C<ev_async_send>,
-reset when the event loop detects that).
+compressed into a single callback invocation (another way to look at
+this is that C<ev_async> watchers are level-triggered: they are set on
+C<ev_async_send>, reset when the event loop detects that).
 
-This call incurs the overhead of a system call only once per event loop
-iteration, so while the overhead might be noticeable, it doesn't apply to
-repeated calls to C<ev_async_send> for the same event loop.
+This call incurs the overhead of at most one extra system call per event
+loop iteration, if the event loop is blocked, and no syscall at all if
+the event loop (or your program) is processing events. That means that
+repeated calls are basically free (there is no need to avoid calls for
+performance reasons) and that the overhead becomes smaller (typically
+zero) under load.
 
 =item bool = ev_async_pending (ev_async *)
 
@@ -4371,10 +4371,11 @@ indicate GNU/Linux + Glibc 2.4 or newer, otherwise disabled.
 =item EV_ATOMIC_T
 
 Libev requires an integer type (suitable for storing C<0> or C<1>) whose
-access is atomic with respect to other threads or signal contexts. No such
-type is easily found in the C language, so you can provide your own type
-that you know is safe for your purposes. It is used both for signal handler "locking"
-as well as for signal and thread safety in C<ev_async> watchers.
+access is atomic and serialised with respect to other threads or signal
+contexts. No such type is easily found in the C language, so you can
+provide your own type that you know is safe for your purposes. It is used
+both for signal handler "locking" as well as for signal and thread safety
+in C<ev_async> watchers.
 
 In the absence of this define, libev will use C<sig_atomic_t volatile>
 (from F<signal.h>), which is usually good enough on most platforms.
@@ -5122,8 +5123,9 @@ watchers becomes O(1) with respect to priority handling.
 =item Processing signals: O(max_signal_number)
 
 Sending involves a system call I<iff> there were no other C<ev_async_send>
-calls in the current loop iteration. Checking for async and signal events
-involves iterating over all running async watchers or all signal numbers.
+calls in the current loop iteration and the loop is currently
+blocked. Checking for async and signal events involves iterating over all
+running async watchers or all signal numbers.
 
 =back
 
diff --git a/ev_vars.h b/ev_vars.h
index 0f1fecf..db132cb 100644
--- a/ev_vars.h
+++ b/ev_vars.h
@@ -73,6 +73,8 @@ VARx(int, evfd)
 #endif
 VAR (evpipe, int evpipe [2])
 VARx(ev_io, pipe_w)
+VARx(EV_ATOMIC_T, pipe_write_wanted)
+VARx(EV_ATOMIC_T, pipe_write_skipped)
 
 #if !defined(_WIN32) || EV_GENWRAP
 VARx(pid_t, curpid)
diff --git a/ev_wrap.h b/ev_wrap.h
index 309be2f..2dd8bb4 100644
--- a/ev_wrap.h
+++ b/ev_wrap.h
@@ -25,6 +25,8 @@
 #define evfd ((loop)->evfd)
 #define evpipe ((loop)->evpipe)
 #define pipe_w ((loop)->pipe_w)
+#define pipe_write_wanted ((loop)->pipe_write_wanted)
+#define pipe_write_skipped ((loop)->pipe_write_skipped)
 #define curpid ((loop)->curpid)
 #define postfork ((loop)->postfork)
 #define vec_ri ((loop)->vec_ri)
@@ -122,6 +124,8 @@
 #undef evfd
 #undef evpipe
 #undef pipe_w
+#undef pipe_write_wanted
+#undef pipe_write_skipped
 #undef curpid
 #undef postfork
 #undef vec_ri