--- libev/ev.pod	2009/04/16 07:56:05	1.236
+++ libev/ev.pod	2010/03/22 09:57:01	1.292
@@ -100,13 +100,14 @@
 Libev supports C<select>, C<poll>, the Linux-specific C<epoll>, the
 BSD-specific C<kqueue> and the Solaris-specific event port mechanisms
 for file descriptor events (C<ev_io>), the Linux C<inotify> interface
-(for C<ev_stat>), relative timers (C<ev_timer>), absolute timers
-with customised rescheduling (C<ev_periodic>), synchronous signals
-(C<ev_signal>), process status change events (C<ev_child>), and event
-watchers dealing with the event loop mechanism itself (C<ev_idle>,
-C<ev_embed>, C<ev_prepare> and C<ev_check> watchers) as well as
-file watchers (C<ev_stat>) and even limited support for fork events
-(C<ev_fork>).
+(for C<ev_stat>), Linux eventfd/signalfd (for faster and cleaner
+inter-thread wakeup (C<ev_async>)/signal handling (C<ev_signal>)) relative
+timers (C<ev_timer>), absolute timers with customised rescheduling
+(C<ev_periodic>), synchronous signals (C<ev_signal>), process status
+change events (C<ev_child>), and event watchers dealing with the event
+loop mechanism itself (C<ev_idle>, C<ev_embed>, C<ev_prepare> and
+C<ev_check> watchers) as well as file watchers (C<ev_stat>) and even
+limited support for fork events (C<ev_fork>).
 
 It also is quite fast (see this
 L<benchmark|http://libev.schmorp.de/bench.html> comparing it to libevent
@@ -119,17 +120,17 @@
 more info about various configuration options please have a look at
 B<EMBED> section in this manual. If libev was configured without support
 for multiple event loops, then all functions taking an initial argument of
-name C<loop> (which is always of type C<ev_loop *>) will not have
+name C<loop> (which is always of type C<struct ev_loop *>) will not have
 this argument.
 
 =head2 TIME REPRESENTATION
 
-Libev represents time as a single floating point number, representing the
-(fractional) number of seconds since the (POSIX) epoch (somewhere near
-the beginning of 1970, details are complicated, don't ask). This type is
-called C<ev_tstamp>, which is what you should use too. It usually aliases
-to the C<double> type in C, and when you need to do any calculations on
-it, you should treat it as some floating point value. Unlike the name
+Libev represents time as a single floating point number, representing
+the (fractional) number of seconds since the (POSIX) epoch (somewhere
+near the beginning of 1970, details are complicated, don't ask). This
+type is called C<ev_tstamp>, which is what you should use too. It usually
+aliases to the C<double> type in C. When you need to do any calculations
+on it, you should treat it as some floating point value. Unlike the name
 component C<stamp> might indicate, it is also used for time differences
 throughout libev.
 
@@ -346,9 +347,8 @@
 
 =item C<EVFLAG_FORKCHECK>
 
-Instead of calling C<ev_default_fork> or C<ev_loop_fork> manually after
-a fork, you can also make libev check for a fork in each iteration by
-enabling this flag.
+Instead of calling C<ev_loop_fork> manually after a fork, you can also
+make libev check for a fork in each iteration by enabling this flag.
 
 This works by calling C<getpid ()> on every iteration of the loop,
 and thus this might slow down your event loop if you do a lot of loop
@@ -364,6 +364,26 @@
 This flag setting cannot be overridden or specified in the C<LIBEV_FLAGS>
 environment variable.
 
+=item C<EVFLAG_NOINOTIFY>
+
+When this flag is specified, then libev will not attempt to use the
+I<inotify> API for it's C<ev_stat> watchers. Apart from debugging and
+testing, this flag can be useful to conserve inotify file descriptors, as
+otherwise each loop using C<ev_stat> watchers consumes one inotify handle.
+
+=item C<EVFLAG_SIGNALFD>
+
+When this flag is specified, then libev will attempt to use the
+I<signalfd> API for it's C<ev_signal> (and C<ev_child>) watchers. This API
+delivers signals synchronously, which makes it both faster and might make
+it possible to get the queued signal data. It can also simplify signal
+handling with threads, as long as you properly block signals in your
+threads that are not interested in handling them.
+
+Signalfd will not be used by default as this changes your signal mask, and
+there are a lot of shoddy libraries and programs (glib's threadpool for
+example) that can't properly initialise their signal masks.
+
 =item C<EVBACKEND_SELECT>  (value 1, portable select backend)
 
 This is your standard select(2) backend. Not I<completely> standard, as
@@ -397,6 +417,9 @@
 
 =item C<EVBACKEND_EPOLL>   (value 4, Linux)
 
+Use the linux-specific epoll(7) interface (for both pre- and post-2.6.9
+kernels).
+
 For few fds, this backend is a bit little slower than poll and select,
 but it scales phenomenally better. While poll and select usually scale
 like O(total_fds) where n is the total number of fds (or the highest fd),
@@ -520,9 +543,10 @@
 
 =back
 
-If one or more of these are or'ed into the flags value, then only these
-backends will be tried (in the reverse order as listed here). If none are
-specified, all backends in C<ev_recommended_backends ()> will be tried.
+If one or more of the backend flags are or'ed into the flags value,
+then only these backends will be tried (in the reverse order as listed
+here). If none are specified, all backends in C<ev_recommended_backends
+()> will be tried.
 
 Example: This is the most typical usage.
 
@@ -544,11 +568,9 @@
 =item struct ev_loop *ev_loop_new (unsigned int flags)
 
 Similar to C<ev_default_loop>, but always creates a new event loop that is
-always distinct from the default loop. Unlike the default loop, it cannot
-handle signal and child watchers, and attempts to do so will be greeted by
-undefined behaviour (or a failed assertion if assertions are enabled).
+always distinct from the default loop.
 
-Note that this function I<is> thread-safe, and the recommended way to use
+Note that this function I<is> thread-safe, and one common way to use
 libev with threads is indeed to create one loop per thread, and using the
 default loop in the "main" or "initial" thread.
 
@@ -560,13 +582,12 @@
 
 =item ev_default_destroy ()
 
-Destroys the default loop again (frees all memory and kernel state
-etc.). None of the active event watchers will be stopped in the normal
-sense, so e.g. C<ev_is_active> might still return true. It is your
-responsibility to either stop all watchers cleanly yourself I<before>
-calling this function, or cope with the fact afterwards (which is usually
-the easiest thing, you can just ignore the watchers and/or C<free ()> them
-for example).
+Destroys the default loop (frees all memory and kernel state etc.). None
+of the active event watchers will be stopped in the normal sense, so
+e.g. C<ev_is_active> might still return true. It is your responsibility to
+either stop all watchers cleanly yourself I<before> calling this function,
+or cope with the fact afterwards (which is usually the easiest thing, you
+can just ignore the watchers and/or C<free ()> them for example).
 
 Note that certain global state, such as signal state (and installed signal
 handlers), will not be freed by this function, and related watchers (such
@@ -575,7 +596,7 @@
 In general it is not advisable to call this function except in the
 rare occasion where you really need to free e.g. the signal handling
 pipe fds. If you need dynamically allocated loops it is better to use
-C<ev_loop_new> and C<ev_loop_destroy>).
+C<ev_loop_new> and C<ev_loop_destroy>.
 
 =item ev_loop_destroy (loop)
 
@@ -591,9 +612,15 @@
 sense). You I<must> call it in the child before using any of the libev
 functions, and it will only take effect at the next C<ev_loop> iteration.
 
+Again, you I<have> to call it on I<any> loop that you want to re-use after 
+a fork, I<even if you do not plan to use the loop in the parent>. This is
+because some kernel interfaces *cough* I<kqueue> *cough* do funny things
+during fork.
+
 On the other hand, you only need to call this function in the child
-process if and only if you want to use the event library in the child. If
-you just fork+exec, you don't have to call it at all.
+process if and only if you want to use the event loop in the child. If you
+just fork+exec or create a new loop in the child, you don't have to call
+it at all.
 
 The function itself is quite fast and it's usually not a problem to call
 it just in case after a fork. To make this easy, the function will fit in
@@ -605,23 +632,37 @@
 
 Like C<ev_default_fork>, but acts on an event loop created by
 C<ev_loop_new>. Yes, you have to call this on every allocated event loop
-after fork that you want to re-use in the child, and how you do this is
-entirely your own problem.
+after fork that you want to re-use in the child, and how you keep track of 
+them is entirely your own problem.
 
 =item int ev_is_default_loop (loop)
 
 Returns true when the given loop is, in fact, the default loop, and false
 otherwise.
 
-=item unsigned int ev_loop_count (loop)
+=item unsigned int ev_iteration (loop)
 
-Returns the count of loop iterations for the loop, which is identical to
+Returns the current iteration count for the loop, which is identical to
 the number of times libev did poll for new events. It starts at C<0> and
 happily wraps around with enough iterations.
 
 This value can sometimes be useful as a generation counter of sorts (it
 "ticks" the number of loop iterations), as it roughly corresponds with
-C<ev_prepare> and C<ev_check> calls.
+C<ev_prepare> and C<ev_check> calls - and is incremented between the
+prepare and check phases.
+
+=item unsigned int ev_depth (loop)
+
+Returns the number of times C<ev_loop> was entered minus the number of
+times C<ev_loop> was exited, in other words, the recursion depth.
+
+Outside C<ev_loop>, this number is zero. In a callback, this number is
+C<1>, unless C<ev_loop> was invoked recursively (or from another thread),
+in which case it is higher.
+
+Leaving C<ev_loop> abnormally (setjmp/longjmp, cancelling the thread
+etc.), doesn't count as "exit" - consider this as a hint to avoid such
+ungentleman behaviour unless it's really convenient.
 
 =item unsigned int ev_backend (loop)
 
@@ -646,7 +687,7 @@
 very long time without entering the event loop, updating libev's idea of
 the current time is a good idea.
 
-See also "The special problem of time updates" in the C<ev_timer> section.
+See also L<The special problem of time updates> in the C<ev_timer> section.
 
 =item ev_suspend (loop)
 
@@ -677,8 +718,8 @@
 =item ev_loop (loop, int flags)
 
 Finally, this is it, the event handler. This function usually is called
-after you initialised all your watchers and you want to start handling
-events.
+after you have initialised all your watchers and you want to start
+handling events.
 
 If the flags argument is specified as C<0>, it will not return until
 either no event watchers are active anymore or C<ev_unloop> was called.
@@ -762,9 +803,10 @@
 loop: Every watcher keeps one reference, and as long as the reference
 count is nonzero, C<ev_loop> will not return on its own.
 
-If you have a watcher you never unregister that should not keep C<ev_loop>
-from returning, call ev_unref() after starting, and ev_ref() before
-stopping it.
+This is useful when you have a watcher that you never intend to
+unregister, but that nevertheless should not keep C<ev_loop> from
+returning. In such a case, call C<ev_unref> after starting, and C<ev_ref>
+before stopping it.
 
 As an example, libev itself uses this for its internal signal pipe: It
 is not visible to the libev user and should not keep C<ev_loop> from
@@ -813,7 +855,9 @@
 time collecting I/O events, so you can handle more events per iteration,
 at the cost of increasing latency. Timeouts (both C<ev_periodic> and
 C<ev_timer>) will be not affected. Setting this to a non-null value will
-introduce an additional C<ev_sleep ()> call into most loop iterations.
+introduce an additional C<ev_sleep ()> call into most loop iterations. The
+sleep time ensures that libev will not poll for I/O events more often then
+once per this interval, on average.
 
 Likewise, by setting a higher I<timeout collect interval> you allow libev
 to spend more time collecting timeouts, at the expense of increased
@@ -825,7 +869,11 @@
 interval to a value near C<0.1> or so, which is often enough for
 interactive servers (of course not for games), likewise for timeouts. It
 usually doesn't make much sense to set it to a lower value than C<0.01>,
-as this approaches the timing granularity of most systems.
+as this approaches the timing granularity of most systems. Note that if
+you do transactions with the outside world and you can't increase the
+parallelity, then this setting will limit your transaction rate (if you
+need to poll once per transaction and the I/O collect interval is 0.01,
+then you can't do more than 100 transations per second).
 
 Setting the I<timeout collect interval> can improve the opportunity for
 saving power, as the program will "bundle" timer callback invocations that
@@ -834,6 +882,77 @@
 reduce iterations/wake-ups is to use C<ev_periodic> watchers and make sure
 they fire on, say, one-second boundaries only.
 
+Example: we only need 0.1s timeout granularity, and we wish not to poll
+more often than 100 times per second:
+
+   ev_set_timeout_collect_interval (EV_DEFAULT_UC_ 0.1);
+   ev_set_io_collect_interval (EV_DEFAULT_UC_ 0.01);
+
+=item ev_invoke_pending (loop)
+
+This call will simply invoke all pending watchers while resetting their
+pending state. Normally, C<ev_loop> does this automatically when required,
+but when overriding the invoke callback this call comes handy.
+
+=item int ev_pending_count (loop)
+
+Returns the number of pending watchers - zero indicates that no watchers
+are pending.
+
+=item ev_set_invoke_pending_cb (loop, void (*invoke_pending_cb)(EV_P))
+
+This overrides the invoke pending functionality of the loop: Instead of
+invoking all pending watchers when there are any, C<ev_loop> will call
+this callback instead. This is useful, for example, when you want to
+invoke the actual watchers inside another context (another thread etc.).
+
+If you want to reset the callback, use C<ev_invoke_pending> as new
+callback.
+
+=item ev_set_loop_release_cb (loop, void (*release)(EV_P), void (*acquire)(EV_P))
+
+Sometimes you want to share the same loop between multiple threads. This
+can be done relatively simply by putting mutex_lock/unlock calls around
+each call to a libev function.
+
+However, C<ev_loop> can run an indefinite time, so it is not feasible to
+wait for it to return. One way around this is to wake up the loop via
+C<ev_unloop> and C<av_async_send>, another way is to set these I<release>
+and I<acquire> callbacks on the loop.
+
+When set, then C<release> will be called just before the thread is
+suspended waiting for new events, and C<acquire> is called just
+afterwards.
+
+Ideally, C<release> will just call your mutex_unlock function, and
+C<acquire> will just call the mutex_lock function again.
+
+While event loop modifications are allowed between invocations of
+C<release> and C<acquire> (that's their only purpose after all), no
+modifications done will affect the event loop, i.e. adding watchers will
+have no effect on the set of file descriptors being watched, or the time
+waited. Use an C<ev_async> watcher to wake up C<ev_loop> when you want it
+to take note of any changes you made.
+
+In theory, threads executing C<ev_loop> will be async-cancel safe between
+invocations of C<release> and C<acquire>.
+
+See also the locking example in the C<THREADS> section later in this
+document.
+
+=item ev_set_userdata (loop, void *data)
+
+=item ev_userdata (loop)
+
+Set and retrieve a single C<void *> associated with a loop. When
+C<ev_set_userdata> has never been called, then C<ev_userdata> returns
+C<0.>
+
+These two functions can be used to associate arbitrary data with a loop,
+and are intended solely for the C<invoke_pending_cb>, C<release> and
+C<acquire> callbacks described above, but of course can be (ab-)used for
+any other purpose as well.
+
 =item ev_loop_verify (loop)
 
 This function only does something when C<EV_VERIFY> support has been
@@ -919,7 +1038,7 @@
 The file descriptor in the C<ev_io> watcher has become readable and/or
 writable.
 
-=item C<EV_TIMEOUT>
+=item C<EV_TIMER>
 
 The C<ev_timer> watcher has timed out.
 
@@ -1019,7 +1138,7 @@
    ev_init (&w, my_cb);
    ev_io_set (&w, STDIN_FILENO, EV_READ);
 
-=item C<ev_TYPE_set> (ev_TYPE *, [args])
+=item C<ev_TYPE_set> (ev_TYPE *watcher, [args])
 
 This macro initialises the type-specific parts of a watcher. You need to
 call C<ev_init> at least once before you call this macro, but you can
@@ -1042,7 +1161,7 @@
 
    ev_io_init (&w, my_cb, STDIN_FILENO, EV_READ);
 
-=item C<ev_TYPE_start> (loop *, ev_TYPE *watcher)
+=item C<ev_TYPE_start> (loop, ev_TYPE *watcher)
 
 Starts (activates) the given watcher. Only active watchers will receive
 events. If the watcher is already active nothing will happen.
@@ -1052,7 +1171,7 @@
 
    ev_io_start (EV_DEFAULT_UC, &w);
 
-=item C<ev_TYPE_stop> (loop *, ev_TYPE *watcher)
+=item C<ev_TYPE_stop> (loop, ev_TYPE *watcher)
 
 Stops the given watcher if active, and clears the pending status (whether
 the watcher was active or not).
@@ -1087,7 +1206,7 @@
 Change the callback. You can change the callback at virtually any time
 (modulo threads).
 
-=item ev_set_priority (ev_TYPE *watcher, priority)
+=item ev_set_priority (ev_TYPE *watcher, int priority)
 
 =item int ev_priority (ev_TYPE *watcher)
 
@@ -1129,6 +1248,20 @@
 Sometimes it can be useful to "poll" a watcher instead of waiting for its
 callback to be invoked, which can be accomplished with this function.
 
+=item ev_feed_event (loop, ev_TYPE *watcher, int revents)
+
+Feeds the given event set into the event loop, as if the specified event
+had happened for the specified watcher (which must be a pointer to an
+initialised but not necessarily started event watcher). Obviously you must
+not free the watcher as long as it has pending events.
+
+Stopping the watcher, letting libev invoke it, or calling
+C<ev_clear_pending> will clear the pending event, even if the watcher was
+not started in the first place.
+
+See also C<ev_feed_fd_event> and C<ev_feed_signal_event> for related
+functions that do not need a watcher.
+
 =back
 
 
@@ -1186,14 +1319,14 @@
    static void
    t1_cb (EV_P_ ev_timer *w, int revents)
    {
-     struct my_biggy big = (struct my_biggy *
+     struct my_biggy big = (struct my_biggy *)
        (((char *)w) - offsetof (struct my_biggy, t1));
    }
 
    static void
    t2_cb (EV_P_ ev_timer *w, int revents)
    {
-     struct my_biggy big = (struct my_biggy *
+     struct my_biggy big = (struct my_biggy *)
        (((char *)w) - offsetof (struct my_biggy, t2));
    }
 
@@ -1279,7 +1412,7 @@
    }
 
    static void
-   idle-cb (EV_P_ ev_idle *w, int revents)
+   idle_cb (EV_P_ ev_idle *w, int revents)
    {
      // actual processing
      read (STDIN_FILENO, ...);
@@ -1334,7 +1467,9 @@
 
 If you cannot use non-blocking mode, then force the use of a
 known-to-be-good backend (at the time of this writing, this includes only
-C<EVBACKEND_SELECT> and C<EVBACKEND_POLL>).
+C<EVBACKEND_SELECT> and C<EVBACKEND_POLL>). The same applies to file
+descriptors for which non-blocking operation makes no sense (such as
+files) - libev doesn't guarentee any specific behaviour in that case.
 
 Another thing you have to watch out for is that it is quite easy to
 receive "spurious" readiness notifications, that is your callback might
@@ -1409,6 +1544,44 @@
 ignore SIGPIPE (and maybe make sure you log the exit status of your daemon
 somewhere, as that would have given you a big clue).
 
+=head3 The special problem of accept()ing when you can't
+
+Many implementations of the POSIX C<accept> function (for example,
+found in post-2004 Linux) have the peculiar behaviour of not removing a
+connection from the pending queue in all error cases.
+
+For example, larger servers often run out of file descriptors (because
+of resource limits), causing C<accept> to fail with C<ENFILE> but not
+rejecting the connection, leading to libev signalling readiness on
+the next iteration again (the connection still exists after all), and
+typically causing the program to loop at 100% CPU usage.
+
+Unfortunately, the set of errors that cause this issue differs between
+operating systems, there is usually little the app can do to remedy the
+situation, and no known thread-safe method of removing the connection to
+cope with overload is known (to me).
+
+One of the easiest ways to handle this situation is to just ignore it
+- when the program encounters an overload, it will just loop until the
+situation is over. While this is a form of busy waiting, no OS offers an
+event-based way to handle this situation, so it's the best one can do.
+
+A better way to handle the situation is to log any errors other than
+C<EAGAIN> and C<EWOULDBLOCK>, making sure not to flood the log with such
+messages, and continue as usual, which at least gives the user an idea of
+what could be wrong ("raise the ulimit!"). For extra points one could stop
+the C<ev_io> watcher on the listening fd "for a while", which reduces CPU
+usage.
+
+If your program is single-threaded, then you could also keep a dummy file
+descriptor for overload situations (e.g. by opening F</dev/null>), and
+when you run into C<ENFILE> or C<EMFILE>, close it, run C<accept>,
+close that fd, and create a new dummy fd. This will gracefully refuse
+clients under typical overload conditions.
+
+The last way to handle it is to simply log the error and C<exit>, as
+is often done with C<malloc> failures, but this results in an easy
+opportunity for a DoS attack.
 
 =head3 Watcher-Specific Functions
 
@@ -1465,10 +1638,11 @@
 monotonic clock option helps a lot here).
 
 The callback is guaranteed to be invoked only I<after> its timeout has
-passed. If multiple timers become ready during the same loop iteration
-then the ones with earlier time-out values are invoked before ones with
-later time-out values (but this is no longer true when a callback calls
-C<ev_loop> recursively).
+passed (not I<at>, so on systems with very low-resolution clocks this
+might introduce a small delay). If multiple timers become ready during the
+same loop iteration then the ones with earlier time-out values are invoked
+before ones of the same priority with later time-out values (but this is
+no longer true when a callback calls C<ev_loop> recursively).
 
 =head3 Be smart about timeouts
 
@@ -1522,7 +1696,7 @@
 
 At start:
 
-   ev_timer_init (timer, callback);
+   ev_init (timer, callback);
    timer->repeat = 60.;
    ev_timer_again (loop, timer);
 
@@ -1594,9 +1768,9 @@
 to the current time (meaning we just have some activity :), then call the
 callback, which will "do the right thing" and start the timer:
 
-   ev_timer_init (timer, callback);
+   ev_init (timer, callback);
    last_activity = ev_now (loop);
-   callback (loop, timer, EV_TIMEOUT);
+   callback (loop, timer, EV_TIMER);
 
 And when there is some activity, simply store the current time in
 C<last_activity>, no libev calls at all:
@@ -1665,6 +1839,36 @@
 update of the time returned by C<ev_now ()> by calling C<ev_now_update
 ()>.
 
+=head3 The special problems of suspended animation
+
+When you leave the server world it is quite customary to hit machines that
+can suspend/hibernate - what happens to the clocks during such a suspend?
+
+Some quick tests made with a Linux 2.6.28 indicate that a suspend freezes
+all processes, while the clocks (C<times>, C<CLOCK_MONOTONIC>) continue
+to run until the system is suspended, but they will not advance while the
+system is suspended. That means, on resume, it will be as if the program
+was frozen for a few seconds, but the suspend time will not be counted
+towards C<ev_timer> when a monotonic clock source is used. The real time
+clock advanced as expected, but if it is used as sole clocksource, then a
+long suspend would be detected as a time jump by libev, and timers would
+be adjusted accordingly.
+
+I would not be surprised to see different behaviour in different between
+operating systems, OS versions or even different hardware.
+
+The other form of suspend (job control, or sending a SIGSTOP) will see a
+time jump in the monotonic clocks and the realtime clock. If the program
+is suspended for a very long time, and monotonic clock sources are in use,
+then you can expect C<ev_timer>s to expire as the full suspension time
+will be counted towards the timers. When no monotonic clock source is in
+use, then libev will again assume a timejump and adjust accordingly.
+
+It might be beneficial for this latter case to call C<ev_suspend>
+and C<ev_resume> in code that handles C<SIGTSTP>, to at least get
+deterministic behaviour in this case (you can do nothing against
+C<SIGSTOP>).
+
 =head3 Watcher-Specific Functions and Data Members
 
 =over 4
@@ -1700,6 +1904,18 @@
 This sounds a bit complicated, see L<Be smart about timeouts>, above, for a
 usage example.
 
+=item ev_tstamp ev_timer_remaining (loop, ev_timer *)
+
+Returns the remaining time until a timer fires. If the timer is active,
+then this time is relative to the current event loop time, otherwise it's
+the timeout value currently configured.
+
+That is, after an C<ev_timer_set (w, 5, 7)>, C<ev_timer_remaining> returns
+C<5>. When the timer is started and one second passes, C<ev_timer_remaining>
+will return C<4>. When the timer expires and is restarted, it will return
+roughly C<7> (likely slightly less as callback invocation takes some time,
+too), and so on.
+
 =item ev_tstamp repeat [read-write]
 
 The current C<repeat> value. Will be used each time the watcher times out
@@ -1944,22 +2160,55 @@
 will try it's best to deliver signals synchronously, i.e. as part of the
 normal event processing, like any other event.
 
-If you want signals asynchronously, just use C<sigaction> as you would
-do without libev and forget about sharing the signal. You can even use
-C<ev_async> from a signal handler to synchronously wake up an event loop.
+If you want signals to be delivered truly asynchronously, just use
+C<sigaction> as you would do without libev and forget about sharing
+the signal. You can even use C<ev_async> from a signal handler to
+synchronously wake up an event loop.
+
+You can configure as many watchers as you like for the same signal, but
+only within the same loop, i.e. you can watch for C<SIGINT> in your
+default loop and for C<SIGIO> in another loop, but you cannot watch for
+C<SIGINT> in both the default loop and another loop at the same time. At
+the moment, C<SIGCHLD> is permanently tied to the default loop.
 
-You can configure as many watchers as you like per signal. Only when the
-first watcher gets started will libev actually register a signal handler
+When the first watcher gets started will libev actually register something
 with the kernel (thus it coexists with your own signal handlers as long as
-you don't register any with libev for the same signal). Similarly, when
-the last signal watcher for a signal is stopped, libev will reset the
-signal handler to SIG_DFL (regardless of what it was set to before).
+you don't register any with libev for the same signal).
 
 If possible and supported, libev will install its handlers with
-C<SA_RESTART> behaviour enabled, so system calls should not be unduly
-interrupted. If you have a problem with system calls getting interrupted by
-signals you can block all signals in an C<ev_check> watcher and unblock
-them in an C<ev_prepare> watcher.
+C<SA_RESTART> (or equivalent) behaviour enabled, so system calls should
+not be unduly interrupted. If you have a problem with system calls getting
+interrupted by signals you can block all signals in an C<ev_check> watcher
+and unblock them in an C<ev_prepare> watcher.
+
+=head3 The special problem of inheritance over fork/execve/pthread_create
+
+Both the signal mask (C<sigprocmask>) and the signal disposition
+(C<sigaction>) are unspecified after starting a signal watcher (and after
+stopping it again), that is, libev might or might not block the signal,
+and might or might not set or restore the installed signal handler.
+
+While this does not matter for the signal disposition (libev never
+sets signals to C<SIG_IGN>, so handlers will be reset to C<SIG_DFL> on
+C<execve>), this matters for the signal mask: many programs do not expect
+certain signals to be blocked.
+
+This means that before calling C<exec> (from the child) you should reset
+the signal mask to whatever "default" you expect (all clear is a good
+choice usually).
+
+The simplest way to ensure that the signal mask is reset in the child is
+to install a fork handler with C<pthread_atfork> that resets it. That will
+catch fork calls done by libraries (such as the libc) as well.
+
+In current versions of libev, the signal will not be blocked indefinitely
+unless you use the C<signalfd> API (C<EV_SIGNALFD>). While this reduces
+the window of opportunity for problems, it will not go away, as libev
+I<has> to modify the signal mask, at least temporarily.
+
+So I can't stress this enough: I<If you do not reset your signal mask when
+you expect it to be empty, you have a race condition in your code>. This
+is not a libev-specific thing, this is true for most event libraries.
 
 =head3 Watcher-Specific Functions and Data Members
 
@@ -2001,17 +2250,21 @@
 has been forked (which implies it might have already exited), as long
 as the event loop isn't entered (or is continued from a watcher), i.e.,
 forking and then immediately registering a watcher for the child is fine,
-but forking and registering a watcher a few event loop iterations later is
-not.
+but forking and registering a watcher a few event loop iterations later or
+in the next callback invocation is not.
 
 Only the default event loop is capable of handling signals, and therefore
 you can only register child watchers in the default event loop.
 
+Due to some design glitches inside libev, child watchers will always be
+handled at maximum priority (their priority is set to C<EV_MAXPRI> by
+libev)
+
 =head3 Process Interaction
 
 Libev grabs C<SIGCHLD> as soon as the default event loop is
-initialised. This is necessary to guarantee proper behaviour even if
-the first child watcher is started after the child exits. The occurrence
+initialised. This is necessary to guarantee proper behaviour even if the
+first child watcher is started after the child exits. The occurrence
 of C<SIGCHLD> is recorded asynchronously, but child reaping is done
 synchronously as part of the event loop processing. Libev always reaps all
 children, even ones not watched.
@@ -2031,7 +2284,8 @@
 Currently, the child watcher never gets stopped, even when the
 child terminates, so normally one needs to stop the watcher in the
 callback. Future versions of libev might stop the watcher automatically
-when a child exit is detected.
+when a child exit is detected (calling C<ev_child_stop> twice is not a
+problem).
 
 =head3 Watcher-Specific Functions and Data Members
 
@@ -2367,7 +2621,7 @@
 
    ev_idle *idle_watcher = malloc (sizeof (ev_idle));
    ev_idle_init (idle_watcher, idle_cb);
-   ev_idle_start (loop, idle_cb);
+   ev_idle_start (loop, idle_watcher);
 
 
 =head2 C<ev_prepare> and C<ev_check> - customise your event loop!
@@ -2470,7 +2724,7 @@
      adns_beforepoll (ads, fds, &nfd, &timeout, timeval_from (ev_time ()));
 
      /* the callback is illegal, but won't be called as we stop during check */
-     ev_timer_init (&tw, 0, timeout * 1e-3);
+     ev_timer_init (&tw, 0, timeout * 1e-3, 0.);
      ev_timer_start (loop, &tw);
 
      // create one ev_io per pollfd
@@ -2710,6 +2964,39 @@
 C<ev_default_fork> cheats and calls it in the wrong process, the fork
 handlers will be invoked, too, of course.
 
+=head3 The special problem of life after fork - how is it possible?
+
+Most uses of C<fork()> consist of forking, then some simple calls to ste
+up/change the process environment, followed by a call to C<exec()>. This
+sequence should be handled by libev without any problems.
+
+This changes when the application actually wants to do event handling
+in the child, or both parent in child, in effect "continuing" after the
+fork.
+
+The default mode of operation (for libev, with application help to detect
+forks) is to duplicate all the state in the child, as would be expected
+when I<either> the parent I<or> the child process continues.
+
+When both processes want to continue using libev, then this is usually the
+wrong result. In that case, usually one process (typically the parent) is
+supposed to continue with all watchers in place as before, while the other
+process typically wants to start fresh, i.e. without any active watchers.
+
+The cleanest and most efficient way to achieve that with libev is to
+simply create a new event loop, which of course will be "empty", and
+use that for new watchers. This has the advantage of not touching more
+memory than necessary, and thus avoiding the copy-on-write, and the
+disadvantage of having to use multiple event loops (which do not support
+signal watchers).
+
+When this is not possible, or you want to use the default loop for
+other reasons, then in the process that wants to start "fresh", call
+C<ev_default_destroy ()> followed by C<ev_default_loop (...)>. Destroying
+the default loop will "orphan" (not stop) all registered watchers, so you
+have to be careful not to execute code that modifies those watchers. Note
+also that in that case, you have to re-register any signal watchers.
+
 =head3 Watcher-Specific Functions and Data Members
 
 =over 4
@@ -2748,7 +3035,8 @@
 C<ev_async> does not support queueing of data in any way. The reason
 is that the author does not know of a simple (or any) algorithm for a
 multiple-writer-single-reader queue that works in all cases and doesn't
-need elaborate support such as pthreads.
+need elaborate support such as pthreads or unportable memory access
+semantics.
 
 That means that if you want to queue data, you have to provide your own
 queue. But at least I can tell you how to implement locking around your
@@ -2897,9 +3185,9 @@
 started. Otherwise an C<ev_timer> watcher with after = C<timeout> (and
 repeat = 0) will be started. C<0> is a valid timeout.
 
-The callback has the type C<void (*cb)(int revents, void *arg)> and gets
+The callback has the type C<void (*cb)(int revents, void *arg)> and is
 passed an C<revents> set like normal event callbacks (a combination of
-C<EV_ERROR>, C<EV_READ>, C<EV_WRITE> or C<EV_TIMEOUT>) and the C<arg>
+C<EV_ERROR>, C<EV_READ>, C<EV_WRITE> or C<EV_TIMER>) and the C<arg>
 value passed to C<ev_once>. Note that it is possible to receive I<both>
 a timeout and an io event at the same time - you probably should give io
 events precedence.
@@ -2910,24 +3198,18 @@
    {
      if (revents & EV_READ)
        /* stdin might have data for us, joy! */;
-     else if (revents & EV_TIMEOUT)
+     else if (revents & EV_TIMER)
        /* doh, nothing entered */;
    }
 
    ev_once (STDIN_FILENO, EV_READ, 10., stdin_ready, 0);
 
-=item ev_feed_event (struct ev_loop *, watcher *, int revents)
-
-Feeds the given event set into the event loop, as if the specified event
-had happened for the specified watcher (which must be a pointer to an
-initialised but not necessarily started event watcher).
-
-=item ev_feed_fd_event (struct ev_loop *, int fd, int revents)
+=item ev_feed_fd_event (loop, int fd, int revents)
 
 Feed an event on the given fd, as if a file descriptor backend detected
 the given events it.
 
-=item ev_feed_signal_event (struct ev_loop *loop, int signum)
+=item ev_feed_signal_event (loop, int signum)
 
 Feed an event as if the given signal occurred (C<loop> must be the default
 loop!).
@@ -3017,7 +3299,7 @@
 
 =item ev::TYPE::TYPE ()
 
-=item ev::TYPE::TYPE (struct ev_loop *)
+=item ev::TYPE::TYPE (loop)
 
 =item ev::TYPE::~TYPE
 
@@ -3104,7 +3386,7 @@
    static void io_cb (ev::io &w, int revents) { }
    iow.set <io_cb> ();
 
-=item w->set (struct ev_loop *)
+=item w->set (loop)
 
 Associates a different C<struct ev_loop> with this watcher. You can only
 do this when the watcher is inactive (and not pending either).
@@ -3211,6 +3493,12 @@
 Erkki Seppala has written Ocaml bindings for libev, to be found at
 L<http://modeemi.cs.tut.fi/~flux/software/ocaml-ev/>.
 
+=item Lua
+
+Brian Maher has written a partial interface to libev for lua (at the
+time of this writing, only C<ev_io> and C<ev_timer>), to be found at
+L<http://github.com/brimworks/lua-ev>.
+
 =back
 
 
@@ -3373,12 +3661,19 @@
 =head2 PREPROCESSOR SYMBOLS/MACROS
 
 Libev can be configured via a variety of preprocessor symbols you have to
-define before including any of its files. The default in the absence of
-autoconf is documented for every option.
+define before including (or compiling) any of its files. The default in
+the absence of autoconf is documented for every option.
+
+Symbols marked with "(h)" do not change the ABI, and can have different
+values when compiling libev vs. including F<ev.h>, so it is permissible
+to redefine them before including F<ev.h> without breaking compatibility
+to a compiled library. All other symbols change the ABI, which means all
+users of libev and the libev code itself must be compiled with compatible
+settings.
 
 =over 4
 
-=item EV_STANDALONE
+=item EV_STANDALONE (h)
 
 Must always be C<1> if you do not use autoconf configuration, which
 keeps libev from including F<config.h>, and it also defines dummy
@@ -3386,7 +3681,7 @@
 supported). It will also not define any of the structs usually found in
 F<event.h> that are not directly supported by the libev core alone.
 
-In stanbdalone mode, libev will still try to automatically deduce the
+In standalone mode, libev will still try to automatically deduce the
 configuration, but has to be more conservative.
 
 =item EV_USE_MONOTONIC
@@ -3461,7 +3756,7 @@
 it is assumed that all these functions actually work on fds, even
 on win32. Should not be defined on non-win32 platforms.
 
-=item EV_FD_TO_WIN32_HANDLE
+=item EV_FD_TO_WIN32_HANDLE(fd)
 
 If C<EV_SELECT_IS_WINSOCKET> is enabled, then libev needs a way to map
 file descriptors to socket handles. When not defining this symbol (the
@@ -3469,6 +3764,20 @@
 correct. In some cases, programs use their own file descriptor management,
 in which case they can provide this function to map fds to socket handles.
 
+=item EV_WIN32_HANDLE_TO_FD(handle)
+
+If C<EV_SELECT_IS_WINSOCKET> then libev maps handles to file descriptors
+using the standard C<_open_osfhandle> function. For programs implementing
+their own fd to handle mapping, overwriting this function makes it easier
+to do so. This can be done by defining this macro to an appropriate value.
+
+=item EV_WIN32_CLOSE_FD(fd)
+
+If programs implement their own fd to handle mapping on win32, then this
+macro can be used to override the C<close> function, useful to unregister
+file descriptors again. Note that the replacement function has to close
+the underlying OS handle.
+
 =item EV_USE_POLL
 
 If defined to be C<1>, libev will compile in support for the C<poll>(2)
@@ -3524,24 +3833,24 @@
 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.
 
-=item EV_H
+=item EV_H (h)
 
 The name of the F<ev.h> header file used to include it. The default if
 undefined is C<"ev.h"> in F<event.h>, F<ev.c> and F<ev++.h>. This can be
 used to virtually rename the F<ev.h> header file in case of conflicts.
 
-=item EV_CONFIG_H
+=item EV_CONFIG_H (h)
 
 If C<EV_STANDALONE> isn't C<1>, this variable can be used to override
 F<ev.c>'s idea of where to find the F<config.h> file, similarly to
 C<EV_H>, above.
 
-=item EV_EVENT_H
+=item EV_EVENT_H (h)
 
 Similarly to C<EV_H>, this macro can be used to override F<event.c>'s idea
 of how the F<event.h> header can be found, the default is C<"event.h">.
 
-=item EV_PROTOTYPES
+=item EV_PROTOTYPES (h)
 
 If defined to be C<0>, then F<ev.h> will not define any function
 prototypes, but still define all the structs and other symbols. This is
@@ -3573,60 +3882,130 @@
 If your embedding application does not need any priorities, defining these
 both to C<0> will save some memory and CPU.
 
-=item EV_PERIODIC_ENABLE
+=item EV_PERIODIC_ENABLE, EV_IDLE_ENABLE, EV_EMBED_ENABLE, EV_STAT_ENABLE,
+EV_PREPARE_ENABLE, EV_CHECK_ENABLE, EV_FORK_ENABLE, EV_SIGNAL_ENABLE,
+EV_ASYNC_ENABLE, EV_CHILD_ENABLE.
+
+If undefined or defined to be C<1> (and the platform supports it), then
+the respective watcher type is supported. If defined to be C<0>, then it
+is not. Disabling watcher types mainly saves codesize.
 
-If undefined or defined to be C<1>, then periodic timers are supported. If
-defined to be C<0>, then they are not. Disabling them saves a few kB of
-code.
+=item EV_FEATURES
 
-=item EV_IDLE_ENABLE
+If you need to shave off some kilobytes of code at the expense of some
+speed (but with the full API), you can define this symbol to request
+certain subsets of functionality. The default is to enable all features
+that can be enabled on the platform.
+
+A typical way to use this symbol is to define it to C<0> (or to a bitset
+with some broad features you want) and then selectively re-enable
+additional parts you want, for example if you want everything minimal,
+but multiple event loop support, async and child watchers and the poll
+backend, use this:
+
+   #define EV_FEATURES 0
+   #define EV_MULTIPLICITY 1
+   #define EV_USE_POLL 1
+   #define EV_CHILD_ENABLE 1
+   #define EV_ASYNC_ENABLE 1
 
-If undefined or defined to be C<1>, then idle watchers are supported. If
-defined to be C<0>, then they are not. Disabling them saves a few kB of
-code.
+The actual value is a bitset, it can be a combination of the following
+values:
 
-=item EV_EMBED_ENABLE
+=over 4
 
-If undefined or defined to be C<1>, then embed watchers are supported. If
-defined to be C<0>, then they are not. Embed watchers rely on most other
-watcher types, which therefore must not be disabled.
+=item C<1> - faster/larger code
 
-=item EV_STAT_ENABLE
+Use larger code to speed up some operations.
 
-If undefined or defined to be C<1>, then stat watchers are supported. If
-defined to be C<0>, then they are not.
+Currently this is used to override some inlining decisions (enlarging the  roughly
+30% code size on amd64.
 
-=item EV_FORK_ENABLE
+When optimising for size, use of compiler flags such as C<-Os> with
+gcc recommended, as well as C<-DNDEBUG>, as libev contains a number of
+assertions.
 
-If undefined or defined to be C<1>, then fork watchers are supported. If
-defined to be C<0>, then they are not.
+=item C<2> - faster/larger data structures
 
-=item EV_ASYNC_ENABLE
+Replaces the small 2-heap for timer management by a faster 4-heap, larger
+hash table sizes and so on. This will usually further increase codesize
+and can additionally have an effect on the size of data structures at
+runtime.
 
-If undefined or defined to be C<1>, then async watchers are supported. If
-defined to be C<0>, then they are not.
+=item C<4> - full API configuration
 
-=item EV_MINIMAL
+This enables priorities (sets C<EV_MAXPRI>=2 and C<EV_MINPRI>=-2), and
+enables multiplicity (C<EV_MULTIPLICITY>=1).
 
-If you need to shave off some kilobytes of code at the expense of some
-speed, define this symbol to C<1>. Currently this is used to override some
-inlining decisions, saves roughly 30% code size on amd64. It also selects a
-much smaller 2-heap for timer management over the default 4-heap.
+=item C<8> - full API
+
+This enables a lot of the "lesser used" API functions. See C<ev.h> for
+details on which parts of the API are still available without this
+feature, and do not complain if this subset changes over time.
+
+=item C<16> - enable all optional watcher types
+
+Enables all optional watcher types.  If you want to selectively enable
+only some watcher types other than I/O and timers (e.g. prepare,
+embed, async, child...) you can enable them manually by defining
+C<EV_watchertype_ENABLE> to C<1> instead.
+
+=item C<32> - enable all backends
+
+This enables all backends - without this feature, you need to enable at
+least one backend manually (C<EV_USE_SELECT> is a good choice).
+
+=item C<64> - enable OS-specific "helper" APIs
+
+Enable inotify, eventfd, signalfd and similar OS-specific helper APIs by
+default.
+
+=back
+
+Compiling with C<gcc -Os -DEV_STANDALONE -DEV_USE_EPOLL=1 -DEV_FEATURES=0>
+reduces the compiled size of libev from 24.7Kb code/2.8Kb data to 6.5Kb
+code/0.3Kb data on my GNU/Linux amd64 system, while still giving you I/O
+watchers, timers and monotonic clock support.
+
+With an intelligent-enough linker (gcc+binutils are intelligent enough
+when you use C<-Wl,--gc-sections -ffunction-sections>) functions unused by
+your program might be left out as well - a binary starting a timer and an
+I/O watcher then might come out at only 5Kb.
+
+=item EV_AVOID_STDIO
+
+If this is set to C<1> at compiletime, then libev will avoid using stdio
+functions (printf, scanf, perror etc.). This will increase the codesize
+somewhat, but if your program doesn't otherwise depend on stdio and your
+libc allows it, this avoids linking in the stdio library which is quite
+big.
+
+Note that error messages might become less precise when this option is
+enabled.
+
+=item EV_NSIG
+
+The highest supported signal number, +1 (or, the number of
+signals): Normally, libev tries to deduce the maximum number of signals
+automatically, but sometimes this fails, in which case it can be
+specified. Also, using a lower number than detected (C<32> should be
+good for about any system in existance) can save some memory, as libev
+statically allocates some 12-24 bytes per signal number.
 
 =item EV_PID_HASHSIZE
 
 C<ev_child> watchers use a small hash table to distribute workload by
-pid. The default size is C<16> (or C<1> with C<EV_MINIMAL>), usually more
-than enough. If you need to manage thousands of children you might want to
-increase this value (I<must> be a power of two).
+pid. The default size is C<16> (or C<1> with C<EV_FEATURES> disabled),
+usually more than enough. If you need to manage thousands of children you
+might want to increase this value (I<must> be a power of two).
 
 =item EV_INOTIFY_HASHSIZE
 
 C<ev_stat> watchers use a small hash table to distribute workload by
-inotify watch id. The default size is C<16> (or C<1> with C<EV_MINIMAL>),
-usually more than enough. If you need to manage thousands of C<ev_stat>
-watchers you might want to increase this value (I<must> be a power of
-two).
+inotify watch id. The default size is C<16> (or C<1> with C<EV_FEATURES>
+disabled), usually more than enough. If you need to manage thousands of
+C<ev_stat> watchers you might want to increase this value (I<must> be a
+power of two).
 
 =item EV_USE_4HEAP
 
@@ -3635,8 +4014,8 @@
 to C<1>. The 4-heap uses more complicated (longer) code but has noticeably
 faster performance with many (thousands) of watchers.
 
-The default is C<1> unless C<EV_MINIMAL> is set in which case it is C<0>
-(disabled).
+The default is C<1>, unless C<EV_FEATURES> overrides it, in which case it
+will be C<0>.
 
 =item EV_HEAP_CACHE_AT
 
@@ -3647,8 +4026,8 @@
 but avoids random read accesses on heap changes. This improves performance
 noticeably with many (hundreds) of watchers.
 
-The default is C<1> unless C<EV_MINIMAL> is set in which case it is C<0>
-(disabled).
+The default is C<1>, unless C<EV_FEATURES> overrides it, in which case it
+will be C<0>.
 
 =item EV_VERIFY
 
@@ -3660,8 +4039,8 @@
 verification code will be called very frequently, which will slow down
 libev considerably.
 
-The default is C<1>, unless C<EV_MINIMAL> is set, in which case it will be
-C<0>.
+The default is C<1>, unless C<EV_FEATURES> overrides it, in which case it
+will be C<0>.
 
 =item EV_COMMON
 
@@ -3729,15 +4108,14 @@
 The usage in rxvt-unicode is simpler. It has a F<ev_cpp.h> header file
 that everybody includes and which overrides some configure choices:
 
-   #define EV_MINIMAL 1
-   #define EV_USE_POLL 0
-   #define EV_MULTIPLICITY 0
-   #define EV_PERIODIC_ENABLE 0
-   #define EV_STAT_ENABLE 0
-   #define EV_FORK_ENABLE 0
+   #define EV_FEATURES 8
+   #define EV_USE_SELECT 1
+   #define EV_PREPARE_ENABLE 1
+   #define EV_IDLE_ENABLE 1
+   #define EV_SIGNAL_ENABLE 1
+   #define EV_CHILD_ENABLE 1
+   #define EV_USE_STDEXCEPT 0
    #define EV_CONFIG_H <config.h>
-   #define EV_MINPRI 0
-   #define EV_MAXPRI 0
 
    #include "ev++.h"
 
@@ -3807,14 +4185,152 @@
 
 =back
 
+=head4 THREAD LOCKING EXAMPLE
+
+Here is a fictitious example of how to run an event loop in a different
+thread than where callbacks are being invoked and watchers are
+created/added/removed.
+
+For a real-world example, see the C<EV::Loop::Async> perl module,
+which uses exactly this technique (which is suited for many high-level
+languages).
+
+The example uses a pthread mutex to protect the loop data, a condition
+variable to wait for callback invocations, an async watcher to notify the
+event loop thread and an unspecified mechanism to wake up the main thread.
+
+First, you need to associate some data with the event loop:
+
+   typedef struct {
+     mutex_t lock; /* global loop lock */
+     ev_async async_w;
+     thread_t tid;
+     cond_t invoke_cv;
+   } userdata;
+
+   void prepare_loop (EV_P)
+   {
+      // for simplicity, we use a static userdata struct.
+      static userdata u;
+
+      ev_async_init (&u->async_w, async_cb);
+      ev_async_start (EV_A_ &u->async_w);
+
+      pthread_mutex_init (&u->lock, 0);
+      pthread_cond_init (&u->invoke_cv, 0);
+
+      // now associate this with the loop
+      ev_set_userdata (EV_A_ u);
+      ev_set_invoke_pending_cb (EV_A_ l_invoke);
+      ev_set_loop_release_cb (EV_A_ l_release, l_acquire);
+
+      // then create the thread running ev_loop
+      pthread_create (&u->tid, 0, l_run, EV_A);
+   }
+
+The callback for the C<ev_async> watcher does nothing: the watcher is used
+solely to wake up the event loop so it takes notice of any new watchers
+that might have been added:
+
+   static void
+   async_cb (EV_P_ ev_async *w, int revents)
+   {
+      // just used for the side effects
+   }
+
+The C<l_release> and C<l_acquire> callbacks simply unlock/lock the mutex
+protecting the loop data, respectively.
+
+   static void
+   l_release (EV_P)
+   {
+     userdata *u = ev_userdata (EV_A);
+     pthread_mutex_unlock (&u->lock);
+   }
+
+   static void
+   l_acquire (EV_P)
+   {
+     userdata *u = ev_userdata (EV_A);
+     pthread_mutex_lock (&u->lock);
+   }
+
+The event loop thread first acquires the mutex, and then jumps straight
+into C<ev_loop>:
+
+   void *
+   l_run (void *thr_arg)
+   {
+     struct ev_loop *loop = (struct ev_loop *)thr_arg;
+
+     l_acquire (EV_A);
+     pthread_setcanceltype (PTHREAD_CANCEL_ASYNCHRONOUS, 0);
+     ev_loop (EV_A_ 0);
+     l_release (EV_A);
+
+     return 0;
+   }
+
+Instead of invoking all pending watchers, the C<l_invoke> callback will
+signal the main thread via some unspecified mechanism (signals? pipe
+writes? C<Async::Interrupt>?) and then waits until all pending watchers
+have been called (in a while loop because a) spurious wakeups are possible
+and b) skipping inter-thread-communication when there are no pending
+watchers is very beneficial):
+
+   static void
+   l_invoke (EV_P)
+   {
+     userdata *u = ev_userdata (EV_A);
+
+     while (ev_pending_count (EV_A))
+       {
+         wake_up_other_thread_in_some_magic_or_not_so_magic_way ();
+         pthread_cond_wait (&u->invoke_cv, &u->lock);
+       }
+   }
+
+Now, whenever the main thread gets told to invoke pending watchers, it
+will grab the lock, call C<ev_invoke_pending> and then signal the loop
+thread to continue:
+
+   static void
+   real_invoke_pending (EV_P)
+   {
+     userdata *u = ev_userdata (EV_A);
+
+     pthread_mutex_lock (&u->lock);
+     ev_invoke_pending (EV_A);
+     pthread_cond_signal (&u->invoke_cv);
+     pthread_mutex_unlock (&u->lock);
+   }
+
+Whenever you want to start/stop a watcher or do other modifications to an
+event loop, you will now have to lock:
+
+   ev_timer timeout_watcher;
+   userdata *u = ev_userdata (EV_A);
+
+   ev_timer_init (&timeout_watcher, timeout_cb, 5.5, 0.);
+
+   pthread_mutex_lock (&u->lock);
+   ev_timer_start (EV_A_ &timeout_watcher);
+   ev_async_send (EV_A_ &u->async_w);
+   pthread_mutex_unlock (&u->lock);
+
+Note that sending the C<ev_async> watcher is required because otherwise
+an event loop currently blocking in the kernel will have no knowledge
+about the newly added timer. By waking up the loop it will pick up any new
+watchers in the next event loop iteration.
+
 =head3 COROUTINES
 
 Libev is very accommodating to coroutines ("cooperative threads"):
 libev fully supports nesting calls to its functions from different
 coroutines (e.g. you can call C<ev_loop> on the same loop from two
-different coroutines, and switch freely between both coroutines running the
-loop, as long as you don't confuse yourself). The only exception is that
-you must not do this from C<ev_periodic> reschedule callbacks.
+different coroutines, and switch freely between both coroutines running
+the loop, as long as you don't confuse yourself). The only exception is
+that you must not do this from C<ev_periodic> reschedule callbacks.
 
 Care has been taken to ensure that libev does not keep local state inside
 C<ev_loop>, and other calls do not usually allow for coroutine switches as
@@ -3901,6 +4417,9 @@
 There is no supported compilation method available on windows except
 embedding it into other applications.
 
+Sensible signal handling is officially unsupported by Microsoft - libev
+tries its best, but under most conditions, signals will simply not work.
+
 Not a libev limitation but worth mentioning: windows apparently doesn't
 accept large writes: instead of resulting in a partial write, windows will
 either accept everything or return C<ENOBUFS> if the buffer is too large,
@@ -3914,7 +4433,7 @@
 more than a hundred or so sockets, then likely it needs to use a totally
 different implementation for windows, as libev offers the POSIX readiness
 notification model, which cannot be implemented efficiently on windows
-(Microsoft monopoly games).
+(due to Microsoft monopoly games).
 
 A typical way to use libev under windows is to embed it (see the embedding
 section for details) and use the following F<evwrap.h> header file instead
@@ -3960,24 +4479,22 @@
 of C<64> handles (probably owning to the fact that all windows kernels
 can only wait for C<64> things at the same time internally; Microsoft
 recommends spawning a chain of threads and wait for 63 handles and the
-previous thread in each. Great).
+previous thread in each. Sounds great!).
 
 Newer versions support more handles, but you need to define C<FD_SETSIZE>
 to some high number (e.g. C<2048>) before compiling the winsocket select
-call (which might be in libev or elsewhere, for example, perl does its own
-select emulation on windows).
+call (which might be in libev or elsewhere, for example, perl and many
+other interpreters do their own select emulation on windows).
 
 Another limit is the number of file descriptors in the Microsoft runtime
-libraries, which by default is C<64> (there must be a hidden I<64> fetish
-or something like this inside Microsoft). You can increase this by calling
-C<_setmaxstdio>, which can increase this limit to C<2048> (another
-arbitrary limit), but is broken in many versions of the Microsoft runtime
-libraries.
-
-This might get you to about C<512> or C<2048> sockets (depending on
-windows version and/or the phase of the moon). To get more, you need to
-wrap all I/O functions and provide your own fd management, but the cost of
-calling select (O(n²)) will likely make this unworkable.
+libraries, which by default is C<64> (there must be a hidden I<64>
+fetish or something like this inside Microsoft). You can increase this
+by calling C<_setmaxstdio>, which can increase this limit to C<2048>
+(another arbitrary limit), but is broken in many versions of the Microsoft
+runtime libraries. This might get you to about C<512> or C<2048> sockets
+(depending on windows version and/or the phase of the moon). To get more,
+you need to wrap all I/O functions and provide your own fd management, but
+the cost of calling select (O(n²)) will likely make this unworkable.
 
 =back
 
@@ -4030,7 +4547,9 @@
 The type C<double> is used to represent timestamps. It is required to
 have at least 51 bits of mantissa (and 9 bits of exponent), which is good
 enough for at least into the year 4000. This requirement is fulfilled by
-implementations implementing IEEE 754 (basically all existing ones).
+implementations implementing IEEE 754, which is basically all existing
+ones. With IEEE 754 doubles, you get microsecond accuracy until at least
+2200.
 
 =back
 
@@ -4108,6 +4627,45 @@
 =back
 
 
+=head1 PORTING FROM LIBEV 3.X TO 4.X
+
+The major version 4 introduced some minor incompatible changes to the API.
+
+At the moment, the C<ev.h> header file tries to implement superficial
+compatibility, so most programs should still compile. Those might be
+removed in later versions of libev, so better update early than late.
+
+=over 4
+
+=item C<ev_loop_count> renamed to C<ev_iteration>
+
+=item C<ev_loop_depth> renamed to C<ev_depth>
+
+=item C<ev_loop_verify> renamed to C<ev_verify>
+
+Most functions working on C<struct ev_loop> objects don't have an
+C<ev_loop_> prefix, so it was removed. Note that C<ev_loop_fork> is
+still called C<ev_loop_fork> because it would otherwise clash with the
+C<ev_fork> typedef.
+
+=item C<EV_TIMEOUT> renamed to C<EV_TIMER> in C<revents>
+
+This is a simple rename - all other watcher types use their name
+as revents flag, and now C<ev_timer> does, too.
+
+Both C<EV_TIMER> and C<EV_TIMEOUT> symbols were present in 3.x versions
+and continue to be present for the forseeable future, so this is mostly a
+documentation change.
+
+=item C<EV_MINIMAL> mechanism replaced by C<EV_FEATURES>
+
+The preprocessor symbol C<EV_MINIMAL> has been replaced by a different
+mechanism, C<EV_FEATURES>. Programs using C<EV_MINIMAL> usually compile
+and work, but the library code will of course be larger.
+
+=back
+
+
 =head1 GLOSSARY
 
 =over 4
@@ -4138,7 +4696,7 @@
 any other events happening anymore.
 
 In libev, events are represented as single bits (such as C<EV_READ> or
-C<EV_TIMEOUT>).
+C<EV_TIMER>).
 
 =item event library