ViewVC Help

View File | Revision Log | Show Annotations | Download File

/cvs/libev/ev.pod

Comparing libev/ev.pod (file contents):
Revision 1.351 by root, Mon Jan 10 14:24:26 2011 UTC vs.
Revision 1.397 by root, Mon Feb 13 01:52:14 2012 UTC

…		…
58	ev_timer_start (loop, &timeout_watcher);	58	ev_timer_start (loop, &timeout_watcher);
59		59
60	// now wait for events to arrive	60	// now wait for events to arrive
61	ev_run (loop, 0);	61	ev_run (loop, 0);
62		62
63	// unloop was called, so exit	63	// break was called, so exit
64	return 0;	64	return 0;
65	}	65	}
66		66
67	=head1 ABOUT THIS DOCUMENT	67	=head1 ABOUT THIS DOCUMENT
68		68
…		…
174	=item ev_tstamp ev_time ()	174	=item ev_tstamp ev_time ()
175		175
176	Returns the current time as libev would use it. Please note that the	176	Returns the current time as libev would use it. Please note that the
177	C<ev_now> function is usually faster and also often returns the timestamp	177	C<ev_now> function is usually faster and also often returns the timestamp
178	you actually want to know. Also interesting is the combination of	178	you actually want to know. Also interesting is the combination of
179	C<ev_update_now> and C<ev_now>.	179	C<ev_now_update> and C<ev_now>.
180		180
181	=item ev_sleep (ev_tstamp interval)	181	=item ev_sleep (ev_tstamp interval)
182		182
183	Sleep for the given interval: The current thread will be blocked until	183	Sleep for the given interval: The current thread will be blocked
184	either it is interrupted or the given time interval has passed. Basically	184	until either it is interrupted or the given time interval has
		185	passed (approximately - it might return a bit earlier even if not
		186	interrupted). Returns immediately if C<< interval <= 0 >>.
		187
185	this is a sub-second-resolution C<sleep ()>.	188	Basically this is a sub-second-resolution C<sleep ()>.
		189
		190	The range of the C<interval> is limited - libev only guarantees to work
		191	with sleep times of up to one day (C<< interval <= 86400 >>).
186		192
187	=item int ev_version_major ()	193	=item int ev_version_major ()
188		194
189	=item int ev_version_minor ()	195	=item int ev_version_minor ()
190		196
…		…
435	example) that can't properly initialise their signal masks.	441	example) that can't properly initialise their signal masks.
436		442
437	=item C<EVFLAG_NOSIGMASK>	443	=item C<EVFLAG_NOSIGMASK>
438		444
439	When this flag is specified, then libev will avoid to modify the signal	445	When this flag is specified, then libev will avoid to modify the signal
440	mask. Specifically, this means you ahve to make sure signals are unblocked	446	mask. Specifically, this means you have to make sure signals are unblocked
441	when you want to receive them.	447	when you want to receive them.
442		448
443	This behaviour is useful when you want to do your own signal handling, or	449	This behaviour is useful when you want to do your own signal handling, or
444	want to handle signals only in specific threads and want to avoid libev	450	want to handle signals only in specific threads and want to avoid libev
445	unblocking the signals.	451	unblocking the signals.
		452
		453	It's also required by POSIX in a threaded program, as libev calls
		454	C<sigprocmask>, whose behaviour is officially unspecified.
446		455
447	This flag's behaviour will become the default in future versions of libev.	456	This flag's behaviour will become the default in future versions of libev.
448		457
449	=item C<EVBACKEND_SELECT> (value 1, portable select backend)	458	=item C<EVBACKEND_SELECT> (value 1, portable select backend)
450		459
…		…
480	=item C<EVBACKEND_EPOLL> (value 4, Linux)	489	=item C<EVBACKEND_EPOLL> (value 4, Linux)
481		490
482	Use the linux-specific epoll(7) interface (for both pre- and post-2.6.9	491	Use the linux-specific epoll(7) interface (for both pre- and post-2.6.9
483	kernels).	492	kernels).
484		493
485	For few fds, this backend is a bit little slower than poll and select,	494	For few fds, this backend is a bit little slower than poll and select, but
486	but it scales phenomenally better. While poll and select usually scale	495	it scales phenomenally better. While poll and select usually scale like
487	like O(total_fds) where n is the total number of fds (or the highest fd),	496	O(total_fds) where total_fds is the total number of fds (or the highest
488	epoll scales either O(1) or O(active_fds).	497	fd), epoll scales either O(1) or O(active_fds).
489		498
490	The epoll mechanism deserves honorable mention as the most misdesigned	499	The epoll mechanism deserves honorable mention as the most misdesigned
491	of the more advanced event mechanisms: mere annoyances include silently	500	of the more advanced event mechanisms: mere annoyances include silently
492	dropping file descriptors, requiring a system call per change per file	501	dropping file descriptors, requiring a system call per change per file
493	descriptor (and unnecessary guessing of parameters), problems with dup,	502	descriptor (and unnecessary guessing of parameters), problems with dup,
…		…
496	0.1ms) and so on. The biggest issue is fork races, however - if a program	505	0.1ms) and so on. The biggest issue is fork races, however - if a program
497	forks then I<both> parent and child process have to recreate the epoll	506	forks then I<both> parent and child process have to recreate the epoll
498	set, which can take considerable time (one syscall per file descriptor)	507	set, which can take considerable time (one syscall per file descriptor)
499	and is of course hard to detect.	508	and is of course hard to detect.
500		509
501	Epoll is also notoriously buggy - embedding epoll fds I<should> work, but	510	Epoll is also notoriously buggy - embedding epoll fds I<should> work,
502	of course I<doesn't>, and epoll just loves to report events for totally	511	but of course I<doesn't>, and epoll just loves to report events for
503	I<different> file descriptors (even already closed ones, so one cannot	512	totally I<different> file descriptors (even already closed ones, so
504	even remove them from the set) than registered in the set (especially	513	one cannot even remove them from the set) than registered in the set
505	on SMP systems). Libev tries to counter these spurious notifications by	514	(especially on SMP systems). Libev tries to counter these spurious
506	employing an additional generation counter and comparing that against the	515	notifications by employing an additional generation counter and comparing
507	events to filter out spurious ones, recreating the set when required. Last	516	that against the events to filter out spurious ones, recreating the set
		517	when required. Epoll also erroneously rounds down timeouts, but gives you
		518	no way to know when and by how much, so sometimes you have to busy-wait
		519	because epoll returns immediately despite a nonzero timeout. And last
508	not least, it also refuses to work with some file descriptors which work	520	not least, it also refuses to work with some file descriptors which work
509	perfectly fine with C<select> (files, many character devices...).	521	perfectly fine with C<select> (files, many character devices...).
510		522
511	Epoll is truly the train wreck analog among event poll mechanisms.	523	Epoll is truly the train wreck among event poll mechanisms, a frankenpoll,
		524	cobbled together in a hurry, no thought to design or interaction with
		525	others. Oh, the pain, will it ever stop...
512		526
513	While stopping, setting and starting an I/O watcher in the same iteration	527	While stopping, setting and starting an I/O watcher in the same iteration
514	will result in some caching, there is still a system call per such	528	will result in some caching, there is still a system call per such
515	incident (because the same I<file descriptor> could point to a different	529	incident (because the same I<file descriptor> could point to a different
516	I<file description> now), so its best to avoid that. Also, C<dup ()>'ed	530	I<file description> now), so its best to avoid that. Also, C<dup ()>'ed
…		…
592	On the positive side, this backend actually performed fully to	606	On the positive side, this backend actually performed fully to
593	specification in all tests and is fully embeddable, which is a rare feat	607	specification in all tests and is fully embeddable, which is a rare feat
594	among the OS-specific backends (I vastly prefer correctness over speed	608	among the OS-specific backends (I vastly prefer correctness over speed
595	hacks).	609	hacks).
596		610
597	On the negative side, the interface is I<bizarre>, with the event polling	611	On the negative side, the interface is I<bizarre> - so bizarre that
		612	even sun itself gets it wrong in their code examples: The event polling
598	function sometimes returning events to the caller even though an error	613	function sometimes returns events to the caller even though an error
599	occured, but with no indication whether it has done so or not (yes, it's	614	occurred, but with no indication whether it has done so or not (yes, it's
600	even documented that way) - deadly for edge-triggered interfaces, but	615	even documented that way) - deadly for edge-triggered interfaces where you
		616	absolutely have to know whether an event occurred or not because you have
		617	to re-arm the watcher.
		618
601	fortunately libev seems to be able to work around it.	619	Fortunately libev seems to be able to work around these idiocies.
602		620
603	This backend maps C<EV_READ> and C<EV_WRITE> in the same way as	621	This backend maps C<EV_READ> and C<EV_WRITE> in the same way as
604	C<EVBACKEND_POLL>.	622	C<EVBACKEND_POLL>.
605		623
606	=item C<EVBACKEND_ALL>	624	=item C<EVBACKEND_ALL>
…		…
816	This is useful if you are waiting for some external event in conjunction	834	This is useful if you are waiting for some external event in conjunction
817	with something not expressible using other libev watchers (i.e. "roll your	835	with something not expressible using other libev watchers (i.e. "roll your
818	own C<ev_run>"). However, a pair of C<ev_prepare>/C<ev_check> watchers is	836	own C<ev_run>"). However, a pair of C<ev_prepare>/C<ev_check> watchers is
819	usually a better approach for this kind of thing.	837	usually a better approach for this kind of thing.
820		838
821	Here are the gory details of what C<ev_run> does:	839	Here are the gory details of what C<ev_run> does (this is for your
		840	understanding, not a guarantee that things will work exactly like this in
		841	future versions):
822		842
823	- Increment loop depth.	843	- Increment loop depth.
824	- Reset the ev_break status.	844	- Reset the ev_break status.
825	- Before the first iteration, call any pending watchers.	845	- Before the first iteration, call any pending watchers.
826	LOOP:	846	LOOP:
…		…
859	anymore.	879	anymore.
860		880
861	... queue jobs here, make sure they register event watchers as long	881	... queue jobs here, make sure they register event watchers as long
862	... as they still have work to do (even an idle watcher will do..)	882	... as they still have work to do (even an idle watcher will do..)
863	ev_run (my_loop, 0);	883	ev_run (my_loop, 0);
864	... jobs done or somebody called unloop. yeah!	884	... jobs done or somebody called break. yeah!
865		885
866	=item ev_break (loop, how)	886	=item ev_break (loop, how)
867		887
868	Can be used to make a call to C<ev_run> return early (but only after it	888	Can be used to make a call to C<ev_run> return early (but only after it
869	has processed all outstanding events). The C<how> argument must be either	889	has processed all outstanding events). The C<how> argument must be either
…		…
932	overhead for the actual polling but can deliver many events at once.	952	overhead for the actual polling but can deliver many events at once.
933		953
934	By setting a higher I<io collect interval> you allow libev to spend more	954	By setting a higher I<io collect interval> you allow libev to spend more
935	time collecting I/O events, so you can handle more events per iteration,	955	time collecting I/O events, so you can handle more events per iteration,
936	at the cost of increasing latency. Timeouts (both C<ev_periodic> and	956	at the cost of increasing latency. Timeouts (both C<ev_periodic> and
937	C<ev_timer>) will be not affected. Setting this to a non-null value will	957	C<ev_timer>) will not be affected. Setting this to a non-null value will
938	introduce an additional C<ev_sleep ()> call into most loop iterations. The	958	introduce an additional C<ev_sleep ()> call into most loop iterations. The
939	sleep time ensures that libev will not poll for I/O events more often then	959	sleep time ensures that libev will not poll for I/O events more often then
940	once per this interval, on average.	960	once per this interval, on average (as long as the host time resolution is
		961	good enough).
941		962
942	Likewise, by setting a higher I<timeout collect interval> you allow libev	963	Likewise, by setting a higher I<timeout collect interval> you allow libev
943	to spend more time collecting timeouts, at the expense of increased	964	to spend more time collecting timeouts, at the expense of increased
944	latency/jitter/inexactness (the watcher callback will be called	965	latency/jitter/inexactness (the watcher callback will be called
945	later). C<ev_io> watchers will not be affected. Setting this to a non-null	966	later). C<ev_io> watchers will not be affected. Setting this to a non-null
…		…
999	can be done relatively simply by putting mutex_lock/unlock calls around	1020	can be done relatively simply by putting mutex_lock/unlock calls around
1000	each call to a libev function.	1021	each call to a libev function.
1001		1022
1002	However, C<ev_run> can run an indefinite time, so it is not feasible	1023	However, C<ev_run> can run an indefinite time, so it is not feasible
1003	to wait for it to return. One way around this is to wake up the event	1024	to wait for it to return. One way around this is to wake up the event
1004	loop via C<ev_break> and C<av_async_send>, another way is to set these	1025	loop via C<ev_break> and C<ev_async_send>, another way is to set these
1005	I<release> and I<acquire> callbacks on the loop.	1026	I<release> and I<acquire> callbacks on the loop.
1006		1027
1007	When set, then C<release> will be called just before the thread is	1028	When set, then C<release> will be called just before the thread is
1008	suspended waiting for new events, and C<acquire> is called just	1029	suspended waiting for new events, and C<acquire> is called just
1009	afterwards.	1030	afterwards.
…		…
1351	See also C<ev_feed_fd_event> and C<ev_feed_signal_event> for related	1372	See also C<ev_feed_fd_event> and C<ev_feed_signal_event> for related
1352	functions that do not need a watcher.	1373	functions that do not need a watcher.
1353		1374
1354	=back	1375	=back
1355		1376
1356	=head2 ASSOCIATING CUSTOM DATA WITH A WATCHER	1377	See also the L<ASSOCIATING CUSTOM DATA WITH A WATCHER> and L<BUILDING YOUR
1357		1378	OWN COMPOSITE WATCHERS> idioms.
1358	Each watcher has, by default, a member C<void *data> that you can change
1359	and read at any time: libev will completely ignore it. This can be used
1360	to associate arbitrary data with your watcher. If you need more data and
1361	don't want to allocate memory and store a pointer to it in that data
1362	member, you can also "subclass" the watcher type and provide your own
1363	data:
1364
1365	struct my_io
1366	{
1367	ev_io io;
1368	int otherfd;
1369	void *somedata;
1370	struct whatever *mostinteresting;
1371	};
1372
1373	...
1374	struct my_io w;
1375	ev_io_init (&w.io, my_cb, fd, EV_READ);
1376
1377	And since your callback will be called with a pointer to the watcher, you
1378	can cast it back to your own type:
1379
1380	static void my_cb (struct ev_loop *loop, ev_io *w_, int revents)
1381	{
1382	struct my_io *w = (struct my_io *)w_;
1383	...
1384	}
1385
1386	More interesting and less C-conformant ways of casting your callback type
1387	instead have been omitted.
1388
1389	Another common scenario is to use some data structure with multiple
1390	embedded watchers:
1391
1392	struct my_biggy
1393	{
1394	int some_data;
1395	ev_timer t1;
1396	ev_timer t2;
1397	}
1398
1399	In this case getting the pointer to C<my_biggy> is a bit more
1400	complicated: Either you store the address of your C<my_biggy> struct
1401	in the C<data> member of the watcher (for woozies), or you need to use
1402	some pointer arithmetic using C<offsetof> inside your watchers (for real
1403	programmers):
1404
1405	#include <stddef.h>
1406
1407	static void
1408	t1_cb (EV_P_ ev_timer *w, int revents)
1409	{
1410	struct my_biggy big = (struct my_biggy *)
1411	(((char *)w) - offsetof (struct my_biggy, t1));
1412	}
1413
1414	static void
1415	t2_cb (EV_P_ ev_timer *w, int revents)
1416	{
1417	struct my_biggy big = (struct my_biggy *)
1418	(((char *)w) - offsetof (struct my_biggy, t2));
1419	}
1420		1379
1421	=head2 WATCHER STATES	1380	=head2 WATCHER STATES
1422		1381
1423	There are various watcher states mentioned throughout this manual -	1382	There are various watcher states mentioned throughout this manual -
1424	active, pending and so on. In this section these states and the rules to	1383	active, pending and so on. In this section these states and the rules to
…		…
1427		1386
1428	=over 4	1387	=over 4
1429		1388
1430	=item initialiased	1389	=item initialiased
1431		1390
1432	Before a watcher can be registered with the event looop it has to be	1391	Before a watcher can be registered with the event loop it has to be
1433	initialised. This can be done with a call to C<ev_TYPE_init>, or calls to	1392	initialised. This can be done with a call to C<ev_TYPE_init>, or calls to
1434	C<ev_init> followed by the watcher-specific C<ev_TYPE_set> function.	1393	C<ev_init> followed by the watcher-specific C<ev_TYPE_set> function.
1435		1394
1436	In this state it is simply some block of memory that is suitable for use	1395	In this state it is simply some block of memory that is suitable for
1437	in an event loop. It can be moved around, freed, reused etc. at will.	1396	use in an event loop. It can be moved around, freed, reused etc. at
		1397	will - as long as you either keep the memory contents intact, or call
		1398	C<ev_TYPE_init> again.
1438		1399
1439	=item started/running/active	1400	=item started/running/active
1440		1401
1441	Once a watcher has been started with a call to C<ev_TYPE_start> it becomes	1402	Once a watcher has been started with a call to C<ev_TYPE_start> it becomes
1442	property of the event loop, and is actively waiting for events. While in	1403	property of the event loop, and is actively waiting for events. While in
…		…
1470	latter will clear any pending state the watcher might be in, regardless	1431	latter will clear any pending state the watcher might be in, regardless
1471	of whether it was active or not, so stopping a watcher explicitly before	1432	of whether it was active or not, so stopping a watcher explicitly before
1472	freeing it is often a good idea.	1433	freeing it is often a good idea.
1473		1434
1474	While stopped (and not pending) the watcher is essentially in the	1435	While stopped (and not pending) the watcher is essentially in the
1475	initialised state, that is it can be reused, moved, modified in any way	1436	initialised state, that is, it can be reused, moved, modified in any way
1476	you wish.	1437	you wish (but when you trash the memory block, you need to C<ev_TYPE_init>
		1438	it again).
1477		1439
1478	=back	1440	=back
1479		1441
1480	=head2 WATCHER PRIORITY MODELS	1442	=head2 WATCHER PRIORITY MODELS
1481		1443
…		…
1610	In general you can register as many read and/or write event watchers per	1572	In general you can register as many read and/or write event watchers per
1611	fd as you want (as long as you don't confuse yourself). Setting all file	1573	fd as you want (as long as you don't confuse yourself). Setting all file
1612	descriptors to non-blocking mode is also usually a good idea (but not	1574	descriptors to non-blocking mode is also usually a good idea (but not
1613	required if you know what you are doing).	1575	required if you know what you are doing).
1614		1576
1615	If you cannot use non-blocking mode, then force the use of a
1616	known-to-be-good backend (at the time of this writing, this includes only
1617	C<EVBACKEND_SELECT> and C<EVBACKEND_POLL>). The same applies to file
1618	descriptors for which non-blocking operation makes no sense (such as
1619	files) - libev doesn't guarantee any specific behaviour in that case.
1620
1621	Another thing you have to watch out for is that it is quite easy to	1577	Another thing you have to watch out for is that it is quite easy to
1622	receive "spurious" readiness notifications, that is your callback might	1578	receive "spurious" readiness notifications, that is, your callback might
1623	be called with C<EV_READ> but a subsequent C<read>(2) will actually block	1579	be called with C<EV_READ> but a subsequent C<read>(2) will actually block
1624	because there is no data. Not only are some backends known to create a	1580	because there is no data. It is very easy to get into this situation even
1625	lot of those (for example Solaris ports), it is very easy to get into	1581	with a relatively standard program structure. Thus it is best to always
1626	this situation even with a relatively standard program structure. Thus	1582	use non-blocking I/O: An extra C<read>(2) returning C<EAGAIN> is far
1627	it is best to always use non-blocking I/O: An extra C<read>(2) returning
1628	C<EAGAIN> is far preferable to a program hanging until some data arrives.	1583	preferable to a program hanging until some data arrives.
1629		1584
1630	If you cannot run the fd in non-blocking mode (for example you should	1585	If you cannot run the fd in non-blocking mode (for example you should
1631	not play around with an Xlib connection), then you have to separately	1586	not play around with an Xlib connection), then you have to separately
1632	re-test whether a file descriptor is really ready with a known-to-be good	1587	re-test whether a file descriptor is really ready with a known-to-be good
1633	interface such as poll (fortunately in our Xlib example, Xlib already	1588	interface such as poll (fortunately in the case of Xlib, it already does
1634	does this on its own, so its quite safe to use). Some people additionally	1589	this on its own, so its quite safe to use). Some people additionally
1635	use C<SIGALRM> and an interval timer, just to be sure you won't block	1590	use C<SIGALRM> and an interval timer, just to be sure you won't block
1636	indefinitely.	1591	indefinitely.
1637		1592
1638	But really, best use non-blocking mode.	1593	But really, best use non-blocking mode.
1639		1594
…		…
1667		1622
1668	There is no workaround possible except not registering events	1623	There is no workaround possible except not registering events
1669	for potentially C<dup ()>'ed file descriptors, or to resort to	1624	for potentially C<dup ()>'ed file descriptors, or to resort to
1670	C<EVBACKEND_SELECT> or C<EVBACKEND_POLL>.	1625	C<EVBACKEND_SELECT> or C<EVBACKEND_POLL>.
1671		1626
		1627	=head3 The special problem of files
		1628
		1629	Many people try to use C<select> (or libev) on file descriptors
		1630	representing files, and expect it to become ready when their program
		1631	doesn't block on disk accesses (which can take a long time on their own).
		1632
		1633	However, this cannot ever work in the "expected" way - you get a readiness
		1634	notification as soon as the kernel knows whether and how much data is
		1635	there, and in the case of open files, that's always the case, so you
		1636	always get a readiness notification instantly, and your read (or possibly
		1637	write) will still block on the disk I/O.
		1638
		1639	Another way to view it is that in the case of sockets, pipes, character
		1640	devices and so on, there is another party (the sender) that delivers data
		1641	on its own, but in the case of files, there is no such thing: the disk
		1642	will not send data on its own, simply because it doesn't know what you
		1643	wish to read - you would first have to request some data.
		1644
		1645	Since files are typically not-so-well supported by advanced notification
		1646	mechanism, libev tries hard to emulate POSIX behaviour with respect
		1647	to files, even though you should not use it. The reason for this is
		1648	convenience: sometimes you want to watch STDIN or STDOUT, which is
		1649	usually a tty, often a pipe, but also sometimes files or special devices
		1650	(for example, C<epoll> on Linux works with F</dev/random> but not with
		1651	F</dev/urandom>), and even though the file might better be served with
		1652	asynchronous I/O instead of with non-blocking I/O, it is still useful when
		1653	it "just works" instead of freezing.
		1654
		1655	So avoid file descriptors pointing to files when you know it (e.g. use
		1656	libeio), but use them when it is convenient, e.g. for STDIN/STDOUT, or
		1657	when you rarely read from a file instead of from a socket, and want to
		1658	reuse the same code path.
		1659
1672	=head3 The special problem of fork	1660	=head3 The special problem of fork
1673		1661
1674	Some backends (epoll, kqueue) do not support C<fork ()> at all or exhibit	1662	Some backends (epoll, kqueue) do not support C<fork ()> at all or exhibit
1675	useless behaviour. Libev fully supports fork, but needs to be told about	1663	useless behaviour. Libev fully supports fork, but needs to be told about
1676	it in the child.	1664	it in the child if you want to continue to use it in the child.
1677		1665
1678	To support fork in your programs, you either have to call	1666	To support fork in your child processes, you have to call C<ev_loop_fork
1679	C<ev_default_fork ()> or C<ev_loop_fork ()> after a fork in the child,	1667	()> after a fork in the child, enable C<EVFLAG_FORKCHECK>, or resort to
1680	enable C<EVFLAG_FORKCHECK>, or resort to C<EVBACKEND_SELECT> or	1668	C<EVBACKEND_SELECT> or C<EVBACKEND_POLL>.
1681	C<EVBACKEND_POLL>.
1682		1669
1683	=head3 The special problem of SIGPIPE	1670	=head3 The special problem of SIGPIPE
1684		1671
1685	While not really specific to libev, it is easy to forget about C<SIGPIPE>:	1672	While not really specific to libev, it is easy to forget about C<SIGPIPE>:
1686	when writing to a pipe whose other end has been closed, your program gets	1673	when writing to a pipe whose other end has been closed, your program gets
…		…
1784	detecting time jumps is hard, and some inaccuracies are unavoidable (the	1771	detecting time jumps is hard, and some inaccuracies are unavoidable (the
1785	monotonic clock option helps a lot here).	1772	monotonic clock option helps a lot here).
1786		1773
1787	The callback is guaranteed to be invoked only I<after> its timeout has	1774	The callback is guaranteed to be invoked only I<after> its timeout has
1788	passed (not I<at>, so on systems with very low-resolution clocks this	1775	passed (not I<at>, so on systems with very low-resolution clocks this
1789	might introduce a small delay). If multiple timers become ready during the	1776	might introduce a small delay, see "the special problem of being too
		1777	early", below). If multiple timers become ready during the same loop
1790	same loop iteration then the ones with earlier time-out values are invoked	1778	iteration then the ones with earlier time-out values are invoked before
1791	before ones of the same priority with later time-out values (but this is	1779	ones of the same priority with later time-out values (but this is no
1792	no longer true when a callback calls C<ev_run> recursively).	1780	longer true when a callback calls C<ev_run> recursively).
1793		1781
1794	=head3 Be smart about timeouts	1782	=head3 Be smart about timeouts
1795		1783
1796	Many real-world problems involve some kind of timeout, usually for error	1784	Many real-world problems involve some kind of timeout, usually for error
1797	recovery. A typical example is an HTTP request - if the other side hangs,	1785	recovery. A typical example is an HTTP request - if the other side hangs,
…		…
1872		1860
1873	In this case, it would be more efficient to leave the C<ev_timer> alone,	1861	In this case, it would be more efficient to leave the C<ev_timer> alone,
1874	but remember the time of last activity, and check for a real timeout only	1862	but remember the time of last activity, and check for a real timeout only
1875	within the callback:	1863	within the callback:
1876		1864
		1865	ev_tstamp timeout = 60.;
1877	ev_tstamp last_activity; // time of last activity	1866	ev_tstamp last_activity; // time of last activity
		1867	ev_timer timer;
1878		1868
1879	static void	1869	static void
1880	callback (EV_P_ ev_timer *w, int revents)	1870	callback (EV_P_ ev_timer *w, int revents)
1881	{	1871	{
1882	ev_tstamp now = ev_now (EV_A);	1872	// calculate when the timeout would happen
1883	ev_tstamp timeout = last_activity + 60.;	1873	ev_tstamp after = last_activity - ev_now (EV_A) + timeout;
1884		1874
1885	// if last_activity + 60. is older than now, we did time out	1875	// if negative, it means we the timeout already occured
1886	if (timeout < now)	1876	if (after < 0.)
1887	{	1877	{
1888	// timeout occurred, take action	1878	// timeout occurred, take action
1889	}	1879	}
1890	else	1880	else
1891	{	1881	{
1892	// callback was invoked, but there was some activity, re-arm	1882	// callback was invoked, but there was some recent
1893	// the watcher to fire in last_activity + 60, which is	1883	// activity. simply restart the timer to time out
1894	// guaranteed to be in the future, so "again" is positive:	1884	// after "after" seconds, which is the earliest time
1895	w->repeat = timeout - now;	1885	// the timeout can occur.
		1886	ev_timer_set (w, after, 0.);
1896	ev_timer_again (EV_A_ w);	1887	ev_timer_start (EV_A_ w);
1897	}	1888	}
1898	}	1889	}
1899		1890
1900	To summarise the callback: first calculate the real timeout (defined	1891	To summarise the callback: first calculate in how many seconds the
1901	as "60 seconds after the last activity"), then check if that time has	1892	timeout will occur (by calculating the absolute time when it would occur,
1902	been reached, which means something I<did>, in fact, time out. Otherwise	1893	C<last_activity + timeout>, and subtracting the current time, C<ev_now
1903	the callback was invoked too early (C<timeout> is in the future), so	1894	(EV_A)> from that).
1904	re-schedule the timer to fire at that future time, to see if maybe we have
1905	a timeout then.
1906		1895
1907	Note how C<ev_timer_again> is used, taking advantage of the	1896	If this value is negative, then we are already past the timeout, i.e. we
1908	C<ev_timer_again> optimisation when the timer is already running.	1897	timed out, and need to do whatever is needed in this case.
		1898
		1899	Otherwise, we now the earliest time at which the timeout would trigger,
		1900	and simply start the timer with this timeout value.
		1901
		1902	In other words, each time the callback is invoked it will check whether
		1903	the timeout cocured. If not, it will simply reschedule itself to check
		1904	again at the earliest time it could time out. Rinse. Repeat.
1909		1905
1910	This scheme causes more callback invocations (about one every 60 seconds	1906	This scheme causes more callback invocations (about one every 60 seconds
1911	minus half the average time between activity), but virtually no calls to	1907	minus half the average time between activity), but virtually no calls to
1912	libev to change the timeout.	1908	libev to change the timeout.
1913		1909
1914	To start the timer, simply initialise the watcher and set C<last_activity>	1910	To start the machinery, simply initialise the watcher and set
1915	to the current time (meaning we just have some activity :), then call the	1911	C<last_activity> to the current time (meaning there was some activity just
1916	callback, which will "do the right thing" and start the timer:	1912	now), then call the callback, which will "do the right thing" and start
		1913	the timer:
1917		1914
		1915	last_activity = ev_now (EV_A);
1918	ev_init (timer, callback);	1916	ev_init (&timer, callback);
1919	last_activity = ev_now (loop);	1917	callback (EV_A_ &timer, 0);
1920	callback (loop, timer, EV_TIMER);
1921		1918
1922	And when there is some activity, simply store the current time in	1919	When there is some activity, simply store the current time in
1923	C<last_activity>, no libev calls at all:	1920	C<last_activity>, no libev calls at all:
1924		1921
		1922	if (activity detected)
1925	last_activity = ev_now (loop);	1923	last_activity = ev_now (EV_A);
		1924
		1925	When your timeout value changes, then the timeout can be changed by simply
		1926	providing a new value, stopping the timer and calling the callback, which
		1927	will agaion do the right thing (for example, time out immediately :).
		1928
		1929	timeout = new_value;
		1930	ev_timer_stop (EV_A_ &timer);
		1931	callback (EV_A_ &timer, 0);
1926		1932
1927	This technique is slightly more complex, but in most cases where the	1933	This technique is slightly more complex, but in most cases where the
1928	time-out is unlikely to be triggered, much more efficient.	1934	time-out is unlikely to be triggered, much more efficient.
1929
1930	Changing the timeout is trivial as well (if it isn't hard-coded in the
1931	callback :) - just change the timeout and invoke the callback, which will
1932	fix things for you.
1933		1935
1934	=item 4. Wee, just use a double-linked list for your timeouts.	1936	=item 4. Wee, just use a double-linked list for your timeouts.
1935		1937
1936	If there is not one request, but many thousands (millions...), all	1938	If there is not one request, but many thousands (millions...), all
1937	employing some kind of timeout with the same timeout value, then one can	1939	employing some kind of timeout with the same timeout value, then one can
…		…
1964	Method #1 is almost always a bad idea, and buys you nothing. Method #4 is	1966	Method #1 is almost always a bad idea, and buys you nothing. Method #4 is
1965	rather complicated, but extremely efficient, something that really pays	1967	rather complicated, but extremely efficient, something that really pays
1966	off after the first million or so of active timers, i.e. it's usually	1968	off after the first million or so of active timers, i.e. it's usually
1967	overkill :)	1969	overkill :)
1968		1970
		1971	=head3 The special problem of being too early
		1972
		1973	If you ask a timer to call your callback after three seconds, then
		1974	you expect it to be invoked after three seconds - but of course, this
		1975	cannot be guaranteed to infinite precision. Less obviously, it cannot be
		1976	guaranteed to any precision by libev - imagine somebody suspending the
		1977	process with a STOP signal for a few hours for example.
		1978
		1979	So, libev tries to invoke your callback as soon as possible I<after> the
		1980	delay has occurred, but cannot guarantee this.
		1981
		1982	A less obvious failure mode is calling your callback too early: many event
		1983	loops compare timestamps with a "elapsed delay >= requested delay", but
		1984	this can cause your callback to be invoked much earlier than you would
		1985	expect.
		1986
		1987	To see why, imagine a system with a clock that only offers full second
		1988	resolution (think windows if you can't come up with a broken enough OS
		1989	yourself). If you schedule a one-second timer at the time 500.9, then the
		1990	event loop will schedule your timeout to elapse at a system time of 500
		1991	(500.9 truncated to the resolution) + 1, or 501.
		1992
		1993	If an event library looks at the timeout 0.1s later, it will see "501 >=
		1994	501" and invoke the callback 0.1s after it was started, even though a
		1995	one-second delay was requested - this is being "too early", despite best
		1996	intentions.
		1997
		1998	This is the reason why libev will never invoke the callback if the elapsed
		1999	delay equals the requested delay, but only when the elapsed delay is
		2000	larger than the requested delay. In the example above, libev would only invoke
		2001	the callback at system time 502, or 1.1s after the timer was started.
		2002
		2003	So, while libev cannot guarantee that your callback will be invoked
		2004	exactly when requested, it I<can> and I<does> guarantee that the requested
		2005	delay has actually elapsed, or in other words, it always errs on the "too
		2006	late" side of things.
		2007
1969	=head3 The special problem of time updates	2008	=head3 The special problem of time updates
1970		2009
1971	Establishing the current time is a costly operation (it usually takes at	2010	Establishing the current time is a costly operation (it usually takes
1972	least two system calls): EV therefore updates its idea of the current	2011	at least one system call): EV therefore updates its idea of the current
1973	time only before and after C<ev_run> collects new events, which causes a	2012	time only before and after C<ev_run> collects new events, which causes a
1974	growing difference between C<ev_now ()> and C<ev_time ()> when handling	2013	growing difference between C<ev_now ()> and C<ev_time ()> when handling
1975	lots of events in one iteration.	2014	lots of events in one iteration.
1976		2015
1977	The relative timeouts are calculated relative to the C<ev_now ()>	2016	The relative timeouts are calculated relative to the C<ev_now ()>
…		…
1983	ev_timer_set (&timer, after + ev_now () - ev_time (), 0.);	2022	ev_timer_set (&timer, after + ev_now () - ev_time (), 0.);
1984		2023
1985	If the event loop is suspended for a long time, you can also force an	2024	If the event loop is suspended for a long time, you can also force an
1986	update of the time returned by C<ev_now ()> by calling C<ev_now_update	2025	update of the time returned by C<ev_now ()> by calling C<ev_now_update
1987	()>.	2026	()>.
		2027
		2028	=head3 The special problem of unsynchronised clocks
		2029
		2030	Modern systems have a variety of clocks - libev itself uses the normal
		2031	"wall clock" clock and, if available, the monotonic clock (to avoid time
		2032	jumps).
		2033
		2034	Neither of these clocks is synchronised with each other or any other clock
		2035	on the system, so C<ev_time ()> might return a considerably different time
		2036	than C<gettimeofday ()> or C<time ()>. On a GNU/Linux system, for example,
		2037	a call to C<gettimeofday> might return a second count that is one higher
		2038	than a directly following call to C<time>.
		2039
		2040	The moral of this is to only compare libev-related timestamps with
		2041	C<ev_time ()> and C<ev_now ()>, at least if you want better precision than
		2042	a second or so.
		2043
		2044	One more problem arises due to this lack of synchronisation: if libev uses
		2045	the system monotonic clock and you compare timestamps from C<ev_time>
		2046	or C<ev_now> from when you started your timer and when your callback is
		2047	invoked, you will find that sometimes the callback is a bit "early".
		2048
		2049	This is because C<ev_timer>s work in real time, not wall clock time, so
		2050	libev makes sure your callback is not invoked before the delay happened,
		2051	I<measured according to the real time>, not the system clock.
		2052
		2053	If your timeouts are based on a physical timescale (e.g. "time out this
		2054	connection after 100 seconds") then this shouldn't bother you as it is
		2055	exactly the right behaviour.
		2056
		2057	If you want to compare wall clock/system timestamps to your timers, then
		2058	you need to use C<ev_periodic>s, as these are based on the wall clock
		2059	time, where your comparisons will always generate correct results.
1988		2060
1989	=head3 The special problems of suspended animation	2061	=head3 The special problems of suspended animation
1990		2062
1991	When you leave the server world it is quite customary to hit machines that	2063	When you leave the server world it is quite customary to hit machines that
1992	can suspend/hibernate - what happens to the clocks during such a suspend?	2064	can suspend/hibernate - what happens to the clocks during such a suspend?
…		…
2036	keep up with the timer (because it takes longer than those 10 seconds to	2108	keep up with the timer (because it takes longer than those 10 seconds to
2037	do stuff) the timer will not fire more than once per event loop iteration.	2109	do stuff) the timer will not fire more than once per event loop iteration.
2038		2110
2039	=item ev_timer_again (loop, ev_timer *)	2111	=item ev_timer_again (loop, ev_timer *)
2040		2112
2041	This will act as if the timer timed out and restart it again if it is	2113	This will act as if the timer timed out, and restarts it again if it is
2042	repeating. The exact semantics are:	2114	repeating. It basically works like calling C<ev_timer_stop>, updating the
		2115	timeout to the C<repeat> value and calling C<ev_timer_start>.
2043		2116
		2117	The exact semantics are as in the following rules, all of which will be
		2118	applied to the watcher:
		2119
		2120	=over 4
		2121
2044	If the timer is pending, its pending status is cleared.	2122	=item If the timer is pending, the pending status is always cleared.
2045		2123
2046	If the timer is started but non-repeating, stop it (as if it timed out).	2124	=item If the timer is started but non-repeating, stop it (as if it timed
		2125	out, without invoking it).
2047		2126
2048	If the timer is repeating, either start it if necessary (with the	2127	=item If the timer is repeating, make the C<repeat> value the new timeout
2049	C<repeat> value), or reset the running timer to the C<repeat> value.	2128	and start the timer, if necessary.
		2129
		2130	=back
2050		2131
2051	This sounds a bit complicated, see L<Be smart about timeouts>, above, for a	2132	This sounds a bit complicated, see L<Be smart about timeouts>, above, for a
2052	usage example.	2133	usage example.
2053		2134
2054	=item ev_tstamp ev_timer_remaining (loop, ev_timer *)	2135	=item ev_tstamp ev_timer_remaining (loop, ev_timer *)
…		…
2176		2257
2177	Another way to think about it (for the mathematically inclined) is that	2258	Another way to think about it (for the mathematically inclined) is that
2178	C<ev_periodic> will try to run the callback in this mode at the next possible	2259	C<ev_periodic> will try to run the callback in this mode at the next possible
2179	time where C<time = offset (mod interval)>, regardless of any time jumps.	2260	time where C<time = offset (mod interval)>, regardless of any time jumps.
2180		2261
2181	For numerical stability it is preferable that the C<offset> value is near	2262	The C<interval> I<MUST> be positive, and for numerical stability, the
2182	C<ev_now ()> (the current time), but there is no range requirement for	2263	interval value should be higher than C<1/8192> (which is around 100
2183	this value, and in fact is often specified as zero.	2264	microseconds) and C<offset> should be higher than C<0> and should have
		2265	at most a similar magnitude as the current time (say, within a factor of
		2266	ten). Typical values for offset are, in fact, C<0> or something between
		2267	C<0> and C<interval>, which is also the recommended range.
2184		2268
2185	Note also that there is an upper limit to how often a timer can fire (CPU	2269	Note also that there is an upper limit to how often a timer can fire (CPU
2186	speed for example), so if C<interval> is very small then timing stability	2270	speed for example), so if C<interval> is very small then timing stability
2187	will of course deteriorate. Libev itself tries to be exact to be about one	2271	will of course deteriorate. Libev itself tries to be exact to be about one
2188	millisecond (if the OS supports it and the machine is fast enough).	2272	millisecond (if the OS supports it and the machine is fast enough).
…		…
2331	=head3 The special problem of inheritance over fork/execve/pthread_create	2415	=head3 The special problem of inheritance over fork/execve/pthread_create
2332		2416
2333	Both the signal mask (C<sigprocmask>) and the signal disposition	2417	Both the signal mask (C<sigprocmask>) and the signal disposition
2334	(C<sigaction>) are unspecified after starting a signal watcher (and after	2418	(C<sigaction>) are unspecified after starting a signal watcher (and after
2335	stopping it again), that is, libev might or might not block the signal,	2419	stopping it again), that is, libev might or might not block the signal,
2336	and might or might not set or restore the installed signal handler.	2420	and might or might not set or restore the installed signal handler (but
		2421	see C<EVFLAG_NOSIGMASK>).
2337		2422
2338	While this does not matter for the signal disposition (libev never	2423	While this does not matter for the signal disposition (libev never
2339	sets signals to C<SIG_IGN>, so handlers will be reset to C<SIG_DFL> on	2424	sets signals to C<SIG_IGN>, so handlers will be reset to C<SIG_DFL> on
2340	C<execve>), this matters for the signal mask: many programs do not expect	2425	C<execve>), this matters for the signal mask: many programs do not expect
2341	certain signals to be blocked.	2426	certain signals to be blocked.
…		…
3212	atexit (program_exits);	3297	atexit (program_exits);
3213		3298
3214		3299
3215	=head2 C<ev_async> - how to wake up an event loop	3300	=head2 C<ev_async> - how to wake up an event loop
3216		3301
3217	In general, you cannot use an C<ev_run> from multiple threads or other	3302	In general, you cannot use an C<ev_loop> from multiple threads or other
3218	asynchronous sources such as signal handlers (as opposed to multiple event	3303	asynchronous sources such as signal handlers (as opposed to multiple event
3219	loops - those are of course safe to use in different threads).	3304	loops - those are of course safe to use in different threads).
3220		3305
3221	Sometimes, however, you need to wake up an event loop you do not control,	3306	Sometimes, however, you need to wake up an event loop you do not control,
3222	for example because it belongs to another thread. This is what C<ev_async>	3307	for example because it belongs to another thread. This is what C<ev_async>
…		…
3229	C<ev_async_sent> calls). In fact, you could use signal watchers as a kind	3314	C<ev_async_sent> calls). In fact, you could use signal watchers as a kind
3230	of "global async watchers" by using a watcher on an otherwise unused	3315	of "global async watchers" by using a watcher on an otherwise unused
3231	signal, and C<ev_feed_signal> to signal this watcher from another thread,	3316	signal, and C<ev_feed_signal> to signal this watcher from another thread,
3232	even without knowing which loop owns the signal.	3317	even without knowing which loop owns the signal.
3233		3318
3234	Unlike C<ev_signal> watchers, C<ev_async> works with any event loop, not
3235	just the default loop.
3236
3237	=head3 Queueing	3319	=head3 Queueing
3238		3320
3239	C<ev_async> does not support queueing of data in any way. The reason	3321	C<ev_async> does not support queueing of data in any way. The reason
3240	is that the author does not know of a simple (or any) algorithm for a	3322	is that the author does not know of a simple (or any) algorithm for a
3241	multiple-writer-single-reader queue that works in all cases and doesn't	3323	multiple-writer-single-reader queue that works in all cases and doesn't
…		…
3332	trust me.	3414	trust me.
3333		3415
3334	=item ev_async_send (loop, ev_async *)	3416	=item ev_async_send (loop, ev_async *)
3335		3417
3336	Sends/signals/activates the given C<ev_async> watcher, that is, feeds	3418	Sends/signals/activates the given C<ev_async> watcher, that is, feeds
3337	an C<EV_ASYNC> event on the watcher into the event loop. Unlike	3419	an C<EV_ASYNC> event on the watcher into the event loop, and instantly
		3420	returns.
		3421
3338	C<ev_feed_event>, this call is safe to do from other threads, signal or	3422	Unlike C<ev_feed_event>, this call is safe to do from other threads,
3339	similar contexts (see the discussion of C<EV_ATOMIC_T> in the embedding	3423	signal or similar contexts (see the discussion of C<EV_ATOMIC_T> in the
3340	section below on what exactly this means).	3424	embedding section below on what exactly this means).
3341		3425
3342	Note that, as with other watchers in libev, multiple events might get	3426	Note that, as with other watchers in libev, multiple events might get
3343	compressed into a single callback invocation (another way to look at this	3427	compressed into a single callback invocation (another way to look at
3344	is that C<ev_async> watchers are level-triggered, set on C<ev_async_send>,	3428	this is that C<ev_async> watchers are level-triggered: they are set on
3345	reset when the event loop detects that).	3429	C<ev_async_send>, reset when the event loop detects that).
3346		3430
3347	This call incurs the overhead of a system call only once per event loop	3431	This call incurs the overhead of at most one extra system call per event
3348	iteration, so while the overhead might be noticeable, it doesn't apply to	3432	loop iteration, if the event loop is blocked, and no syscall at all if
3349	repeated calls to C<ev_async_send> for the same event loop.	3433	the event loop (or your program) is processing events. That means that
		3434	repeated calls are basically free (there is no need to avoid calls for
		3435	performance reasons) and that the overhead becomes smaller (typically
		3436	zero) under load.
3350		3437
3351	=item bool = ev_async_pending (ev_async *)	3438	=item bool = ev_async_pending (ev_async *)
3352		3439
3353	Returns a non-zero value when C<ev_async_send> has been called on the	3440	Returns a non-zero value when C<ev_async_send> has been called on the
3354	watcher but the event has not yet been processed (or even noted) by the	3441	watcher but the event has not yet been processed (or even noted) by the
…		…
3409	ev_once (STDIN_FILENO, EV_READ, 10., stdin_ready, 0);	3496	ev_once (STDIN_FILENO, EV_READ, 10., stdin_ready, 0);
3410		3497
3411	=item ev_feed_fd_event (loop, int fd, int revents)	3498	=item ev_feed_fd_event (loop, int fd, int revents)
3412		3499
3413	Feed an event on the given fd, as if a file descriptor backend detected	3500	Feed an event on the given fd, as if a file descriptor backend detected
3414	the given events it.	3501	the given events.
3415		3502
3416	=item ev_feed_signal_event (loop, int signum)	3503	=item ev_feed_signal_event (loop, int signum)
3417		3504
3418	Feed an event as if the given signal occurred. See also C<ev_feed_signal>,	3505	Feed an event as if the given signal occurred. See also C<ev_feed_signal>,
3419	which is async-safe.	3506	which is async-safe.
…		…
3425		3512
3426	This section explains some common idioms that are not immediately	3513	This section explains some common idioms that are not immediately
3427	obvious. Note that examples are sprinkled over the whole manual, and this	3514	obvious. Note that examples are sprinkled over the whole manual, and this
3428	section only contains stuff that wouldn't fit anywhere else.	3515	section only contains stuff that wouldn't fit anywhere else.
3429		3516
3430	=over 4	3517	=head2 ASSOCIATING CUSTOM DATA WITH A WATCHER
3431		3518
3432	=item Model/nested event loop invocations and exit conditions.	3519	Each watcher has, by default, a C<void *data> member that you can read
		3520	or modify at any time: libev will completely ignore it. This can be used
		3521	to associate arbitrary data with your watcher. If you need more data and
		3522	don't want to allocate memory separately and store a pointer to it in that
		3523	data member, you can also "subclass" the watcher type and provide your own
		3524	data:
		3525
		3526	struct my_io
		3527	{
		3528	ev_io io;
		3529	int otherfd;
		3530	void *somedata;
		3531	struct whatever *mostinteresting;
		3532	};
		3533
		3534	...
		3535	struct my_io w;
		3536	ev_io_init (&w.io, my_cb, fd, EV_READ);
		3537
		3538	And since your callback will be called with a pointer to the watcher, you
		3539	can cast it back to your own type:
		3540
		3541	static void my_cb (struct ev_loop *loop, ev_io *w_, int revents)
		3542	{
		3543	struct my_io *w = (struct my_io *)w_;
		3544	...
		3545	}
		3546
		3547	More interesting and less C-conformant ways of casting your callback
		3548	function type instead have been omitted.
		3549
		3550	=head2 BUILDING YOUR OWN COMPOSITE WATCHERS
		3551
		3552	Another common scenario is to use some data structure with multiple
		3553	embedded watchers, in effect creating your own watcher that combines
		3554	multiple libev event sources into one "super-watcher":
		3555
		3556	struct my_biggy
		3557	{
		3558	int some_data;
		3559	ev_timer t1;
		3560	ev_timer t2;
		3561	}
		3562
		3563	In this case getting the pointer to C<my_biggy> is a bit more
		3564	complicated: Either you store the address of your C<my_biggy> struct in
		3565	the C<data> member of the watcher (for woozies or C++ coders), or you need
		3566	to use some pointer arithmetic using C<offsetof> inside your watchers (for
		3567	real programmers):
		3568
		3569	#include <stddef.h>
		3570
		3571	static void
		3572	t1_cb (EV_P_ ev_timer *w, int revents)
		3573	{
		3574	struct my_biggy big = (struct my_biggy *)
		3575	(((char *)w) - offsetof (struct my_biggy, t1));
		3576	}
		3577
		3578	static void
		3579	t2_cb (EV_P_ ev_timer *w, int revents)
		3580	{
		3581	struct my_biggy big = (struct my_biggy *)
		3582	(((char *)w) - offsetof (struct my_biggy, t2));
		3583	}
		3584
		3585	=head2 AVOIDING FINISHING BEFORE RETURNING
		3586
		3587	Often you have structures like this in event-based programs:
		3588
		3589	callback ()
		3590	{
		3591	free (request);
		3592	}
		3593
		3594	request = start_new_request (..., callback);
		3595
		3596	The intent is to start some "lengthy" operation. The C<request> could be
		3597	used to cancel the operation, or do other things with it.
		3598
		3599	It's not uncommon to have code paths in C<start_new_request> that
		3600	immediately invoke the callback, for example, to report errors. Or you add
		3601	some caching layer that finds that it can skip the lengthy aspects of the
		3602	operation and simply invoke the callback with the result.
		3603
		3604	The problem here is that this will happen I<before> C<start_new_request>
		3605	has returned, so C<request> is not set.
		3606
		3607	Even if you pass the request by some safer means to the callback, you
		3608	might want to do something to the request after starting it, such as
		3609	canceling it, which probably isn't working so well when the callback has
		3610	already been invoked.
		3611
		3612	A common way around all these issues is to make sure that
		3613	C<start_new_request> I<always> returns before the callback is invoked. If
		3614	C<start_new_request> immediately knows the result, it can artificially
		3615	delay invoking the callback by e.g. using a C<prepare> or C<idle> watcher
		3616	for example, or more sneakily, by reusing an existing (stopped) watcher
		3617	and pushing it into the pending queue:
		3618
		3619	ev_set_cb (watcher, callback);
		3620	ev_feed_event (EV_A_ watcher, 0);
		3621
		3622	This way, C<start_new_request> can safely return before the callback is
		3623	invoked, while not delaying callback invocation too much.
		3624
		3625	=head2 MODEL/NESTED EVENT LOOP INVOCATIONS AND EXIT CONDITIONS
3433		3626
3434	Often (especially in GUI toolkits) there are places where you have	3627	Often (especially in GUI toolkits) there are places where you have
3435	I<modal> interaction, which is most easily implemented by recursively	3628	I<modal> interaction, which is most easily implemented by recursively
3436	invoking C<ev_run>.	3629	invoking C<ev_run>.
3437		3630
…		…
3449	int exit_main_loop = 0;	3642	int exit_main_loop = 0;
3450		3643
3451	while (!exit_main_loop)	3644	while (!exit_main_loop)
3452	ev_run (EV_DEFAULT_ EVRUN_ONCE);	3645	ev_run (EV_DEFAULT_ EVRUN_ONCE);
3453		3646
3454	// in a model watcher	3647	// in a modal watcher
3455	int exit_nested_loop = 0;	3648	int exit_nested_loop = 0;
3456		3649
3457	while (!exit_nested_loop)	3650	while (!exit_nested_loop)
3458	ev_run (EV_A_ EVRUN_ONCE);	3651	ev_run (EV_A_ EVRUN_ONCE);
3459		3652
…		…
3466	exit_main_loop = 1;	3659	exit_main_loop = 1;
3467		3660
3468	// exit both	3661	// exit both
3469	exit_main_loop = exit_nested_loop = 1;	3662	exit_main_loop = exit_nested_loop = 1;
3470		3663
3471	=back	3664	=head2 THREAD LOCKING EXAMPLE
		3665
		3666	Here is a fictitious example of how to run an event loop in a different
		3667	thread from where callbacks are being invoked and watchers are
		3668	created/added/removed.
		3669
		3670	For a real-world example, see the C<EV::Loop::Async> perl module,
		3671	which uses exactly this technique (which is suited for many high-level
		3672	languages).
		3673
		3674	The example uses a pthread mutex to protect the loop data, a condition
		3675	variable to wait for callback invocations, an async watcher to notify the
		3676	event loop thread and an unspecified mechanism to wake up the main thread.
		3677
		3678	First, you need to associate some data with the event loop:
		3679
		3680	typedef struct {
		3681	mutex_t lock; /* global loop lock */
		3682	ev_async async_w;
		3683	thread_t tid;
		3684	cond_t invoke_cv;
		3685	} userdata;
		3686
		3687	void prepare_loop (EV_P)
		3688	{
		3689	// for simplicity, we use a static userdata struct.
		3690	static userdata u;
		3691
		3692	ev_async_init (&u->async_w, async_cb);
		3693	ev_async_start (EV_A_ &u->async_w);
		3694
		3695	pthread_mutex_init (&u->lock, 0);
		3696	pthread_cond_init (&u->invoke_cv, 0);
		3697
		3698	// now associate this with the loop
		3699	ev_set_userdata (EV_A_ u);
		3700	ev_set_invoke_pending_cb (EV_A_ l_invoke);
		3701	ev_set_loop_release_cb (EV_A_ l_release, l_acquire);
		3702
		3703	// then create the thread running ev_run
		3704	pthread_create (&u->tid, 0, l_run, EV_A);
		3705	}
		3706
		3707	The callback for the C<ev_async> watcher does nothing: the watcher is used
		3708	solely to wake up the event loop so it takes notice of any new watchers
		3709	that might have been added:
		3710
		3711	static void
		3712	async_cb (EV_P_ ev_async *w, int revents)
		3713	{
		3714	// just used for the side effects
		3715	}
		3716
		3717	The C<l_release> and C<l_acquire> callbacks simply unlock/lock the mutex
		3718	protecting the loop data, respectively.
		3719
		3720	static void
		3721	l_release (EV_P)
		3722	{
		3723	userdata *u = ev_userdata (EV_A);
		3724	pthread_mutex_unlock (&u->lock);
		3725	}
		3726
		3727	static void
		3728	l_acquire (EV_P)
		3729	{
		3730	userdata *u = ev_userdata (EV_A);
		3731	pthread_mutex_lock (&u->lock);
		3732	}
		3733
		3734	The event loop thread first acquires the mutex, and then jumps straight
		3735	into C<ev_run>:
		3736
		3737	void *
		3738	l_run (void *thr_arg)
		3739	{
		3740	struct ev_loop *loop = (struct ev_loop *)thr_arg;
		3741
		3742	l_acquire (EV_A);
		3743	pthread_setcanceltype (PTHREAD_CANCEL_ASYNCHRONOUS, 0);
		3744	ev_run (EV_A_ 0);
		3745	l_release (EV_A);
		3746
		3747	return 0;
		3748	}
		3749
		3750	Instead of invoking all pending watchers, the C<l_invoke> callback will
		3751	signal the main thread via some unspecified mechanism (signals? pipe
		3752	writes? C<Async::Interrupt>?) and then waits until all pending watchers
		3753	have been called (in a while loop because a) spurious wakeups are possible
		3754	and b) skipping inter-thread-communication when there are no pending
		3755	watchers is very beneficial):
		3756
		3757	static void
		3758	l_invoke (EV_P)
		3759	{
		3760	userdata *u = ev_userdata (EV_A);
		3761
		3762	while (ev_pending_count (EV_A))
		3763	{
		3764	wake_up_other_thread_in_some_magic_or_not_so_magic_way ();
		3765	pthread_cond_wait (&u->invoke_cv, &u->lock);
		3766	}
		3767	}
		3768
		3769	Now, whenever the main thread gets told to invoke pending watchers, it
		3770	will grab the lock, call C<ev_invoke_pending> and then signal the loop
		3771	thread to continue:
		3772
		3773	static void
		3774	real_invoke_pending (EV_P)
		3775	{
		3776	userdata *u = ev_userdata (EV_A);
		3777
		3778	pthread_mutex_lock (&u->lock);
		3779	ev_invoke_pending (EV_A);
		3780	pthread_cond_signal (&u->invoke_cv);
		3781	pthread_mutex_unlock (&u->lock);
		3782	}
		3783
		3784	Whenever you want to start/stop a watcher or do other modifications to an
		3785	event loop, you will now have to lock:
		3786
		3787	ev_timer timeout_watcher;
		3788	userdata *u = ev_userdata (EV_A);
		3789
		3790	ev_timer_init (&timeout_watcher, timeout_cb, 5.5, 0.);
		3791
		3792	pthread_mutex_lock (&u->lock);
		3793	ev_timer_start (EV_A_ &timeout_watcher);
		3794	ev_async_send (EV_A_ &u->async_w);
		3795	pthread_mutex_unlock (&u->lock);
		3796
		3797	Note that sending the C<ev_async> watcher is required because otherwise
		3798	an event loop currently blocking in the kernel will have no knowledge
		3799	about the newly added timer. By waking up the loop it will pick up any new
		3800	watchers in the next event loop iteration.
		3801
		3802	=head2 THREADS, COROUTINES, CONTINUATIONS, QUEUES... INSTEAD OF CALLBACKS
		3803
		3804	While the overhead of a callback that e.g. schedules a thread is small, it
		3805	is still an overhead. If you embed libev, and your main usage is with some
		3806	kind of threads or coroutines, you might want to customise libev so that
		3807	doesn't need callbacks anymore.
		3808
		3809	Imagine you have coroutines that you can switch to using a function
		3810	C<switch_to (coro)>, that libev runs in a coroutine called C<libev_coro>
		3811	and that due to some magic, the currently active coroutine is stored in a
		3812	global called C<current_coro>. Then you can build your own "wait for libev
		3813	event" primitive by changing C<EV_CB_DECLARE> and C<EV_CB_INVOKE> (note
		3814	the differing C<;> conventions):
		3815
		3816	#define EV_CB_DECLARE(type) struct my_coro *cb;
		3817	#define EV_CB_INVOKE(watcher) switch_to ((watcher)->cb)
		3818
		3819	That means instead of having a C callback function, you store the
		3820	coroutine to switch to in each watcher, and instead of having libev call
		3821	your callback, you instead have it switch to that coroutine.
		3822
		3823	A coroutine might now wait for an event with a function called
		3824	C<wait_for_event>. (the watcher needs to be started, as always, but it doesn't
		3825	matter when, or whether the watcher is active or not when this function is
		3826	called):
		3827
		3828	void
		3829	wait_for_event (ev_watcher *w)
		3830	{
		3831	ev_cb_set (w) = current_coro;
		3832	switch_to (libev_coro);
		3833	}
		3834
		3835	That basically suspends the coroutine inside C<wait_for_event> and
		3836	continues the libev coroutine, which, when appropriate, switches back to
		3837	this or any other coroutine.
		3838
		3839	You can do similar tricks if you have, say, threads with an event queue -
		3840	instead of storing a coroutine, you store the queue object and instead of
		3841	switching to a coroutine, you push the watcher onto the queue and notify
		3842	any waiters.
		3843
		3844	To embed libev, see L<EMBEDDING>, but in short, it's easiest to create two
		3845	files, F<my_ev.h> and F<my_ev.c> that include the respective libev files:
		3846
		3847	// my_ev.h
		3848	#define EV_CB_DECLARE(type) struct my_coro *cb;
		3849	#define EV_CB_INVOKE(watcher) switch_to ((watcher)->cb);
		3850	#include "../libev/ev.h"
		3851
		3852	// my_ev.c
		3853	#define EV_H "my_ev.h"
		3854	#include "../libev/ev.c"
		3855
		3856	And then use F<my_ev.h> when you would normally use F<ev.h>, and compile
		3857	F<my_ev.c> into your project. When properly specifying include paths, you
		3858	can even use F<ev.h> as header file name directly.
3472		3859
3473		3860
3474	=head1 LIBEVENT EMULATION	3861	=head1 LIBEVENT EMULATION
3475		3862
3476	Libev offers a compatibility emulation layer for libevent. It cannot	3863	Libev offers a compatibility emulation layer for libevent. It cannot
…		…
3530	with C<operator ()> can be used as callbacks. Other types should be easy	3917	with C<operator ()> can be used as callbacks. Other types should be easy
3531	to add as long as they only need one additional pointer for context. If	3918	to add as long as they only need one additional pointer for context. If
3532	you need support for other types of functors please contact the author	3919	you need support for other types of functors please contact the author
3533	(preferably after implementing it).	3920	(preferably after implementing it).
3534		3921
		3922	For all this to work, your C++ compiler either has to use the same calling
		3923	conventions as your C compiler (for static member functions), or you have
		3924	to embed libev and compile libev itself as C++.
		3925
3535	Here is a list of things available in the C<ev> namespace:	3926	Here is a list of things available in the C<ev> namespace:
3536		3927
3537	=over 4	3928	=over 4
3538		3929
3539	=item C<ev::READ>, C<ev::WRITE> etc.	3930	=item C<ev::READ>, C<ev::WRITE> etc.
…		…
3548	=item C<ev::io>, C<ev::timer>, C<ev::periodic>, C<ev::idle>, C<ev::sig> etc.	3939	=item C<ev::io>, C<ev::timer>, C<ev::periodic>, C<ev::idle>, C<ev::sig> etc.
3549		3940
3550	For each C<ev_TYPE> watcher in F<ev.h> there is a corresponding class of	3941	For each C<ev_TYPE> watcher in F<ev.h> there is a corresponding class of
3551	the same name in the C<ev> namespace, with the exception of C<ev_signal>	3942	the same name in the C<ev> namespace, with the exception of C<ev_signal>
3552	which is called C<ev::sig> to avoid clashes with the C<signal> macro	3943	which is called C<ev::sig> to avoid clashes with the C<signal> macro
3553	defines by many implementations.	3944	defined by many implementations.
3554		3945
3555	All of those classes have these methods:	3946	All of those classes have these methods:
3556		3947
3557	=over 4	3948	=over 4
3558		3949
…		…
3691	watchers in the constructor.	4082	watchers in the constructor.
3692		4083
3693	class myclass	4084	class myclass
3694	{	4085	{
3695	ev::io io ; void io_cb (ev::io &w, int revents);	4086	ev::io io ; void io_cb (ev::io &w, int revents);
3696	ev::io2 io2 ; void io2_cb (ev::io &w, int revents);	4087	ev::io io2 ; void io2_cb (ev::io &w, int revents);
3697	ev::idle idle; void idle_cb (ev::idle &w, int revents);	4088	ev::idle idle; void idle_cb (ev::idle &w, int revents);
3698		4089
3699	myclass (int fd)	4090	myclass (int fd)
3700	{	4091	{
3701	io .set <myclass, &myclass::io_cb > (this);	4092	io .set <myclass, &myclass::io_cb > (this);
…		…
3752	L<http://hackage.haskell.org/cgi-bin/hackage-scripts/package/hlibev>.	4143	L<http://hackage.haskell.org/cgi-bin/hackage-scripts/package/hlibev>.
3753		4144
3754	=item D	4145	=item D
3755		4146
3756	Leandro Lucarella has written a D language binding (F<ev.d>) for libev, to	4147	Leandro Lucarella has written a D language binding (F<ev.d>) for libev, to
3757	be found at L<http://proj.llucax.com.ar/wiki/evd>.	4148	be found at L<http://www.llucax.com.ar/proj/ev.d/index.html>.
3758		4149
3759	=item Ocaml	4150	=item Ocaml
3760		4151
3761	Erkki Seppala has written Ocaml bindings for libev, to be found at	4152	Erkki Seppala has written Ocaml bindings for libev, to be found at
3762	L<http://modeemi.cs.tut.fi/~flux/software/ocaml-ev/>.	4153	L<http://modeemi.cs.tut.fi/~flux/software/ocaml-ev/>.
…		…
3810	suitable for use with C<EV_A>.	4201	suitable for use with C<EV_A>.
3811		4202
3812	=item C<EV_DEFAULT>, C<EV_DEFAULT_>	4203	=item C<EV_DEFAULT>, C<EV_DEFAULT_>
3813		4204
3814	Similar to the other two macros, this gives you the value of the default	4205	Similar to the other two macros, this gives you the value of the default
3815	loop, if multiple loops are supported ("ev loop default").	4206	loop, if multiple loops are supported ("ev loop default"). The default loop
		4207	will be initialised if it isn't already initialised.
		4208
		4209	For non-multiplicity builds, these macros do nothing, so you always have
		4210	to initialise the loop somewhere.
3816		4211
3817	=item C<EV_DEFAULT_UC>, C<EV_DEFAULT_UC_>	4212	=item C<EV_DEFAULT_UC>, C<EV_DEFAULT_UC_>
3818		4213
3819	Usage identical to C<EV_DEFAULT> and C<EV_DEFAULT_>, but requires that the	4214	Usage identical to C<EV_DEFAULT> and C<EV_DEFAULT_>, but requires that the
3820	default loop has been initialised (C<UC> == unchecked). Their behaviour	4215	default loop has been initialised (C<UC> == unchecked). Their behaviour
…		…
3965	supported). It will also not define any of the structs usually found in	4360	supported). It will also not define any of the structs usually found in
3966	F<event.h> that are not directly supported by the libev core alone.	4361	F<event.h> that are not directly supported by the libev core alone.
3967		4362
3968	In standalone mode, libev will still try to automatically deduce the	4363	In standalone mode, libev will still try to automatically deduce the
3969	configuration, but has to be more conservative.	4364	configuration, but has to be more conservative.
		4365
		4366	=item EV_USE_FLOOR
		4367
		4368	If defined to be C<1>, libev will use the C<floor ()> function for its
		4369	periodic reschedule calculations, otherwise libev will fall back on a
		4370	portable (slower) implementation. If you enable this, you usually have to
		4371	link against libm or something equivalent. Enabling this when the C<floor>
		4372	function is not available will fail, so the safe default is to not enable
		4373	this.
3970		4374
3971	=item EV_USE_MONOTONIC	4375	=item EV_USE_MONOTONIC
3972		4376
3973	If defined to be C<1>, libev will try to detect the availability of the	4377	If defined to be C<1>, libev will try to detect the availability of the
3974	monotonic clock option at both compile time and runtime. Otherwise no	4378	monotonic clock option at both compile time and runtime. Otherwise no
…		…
4104	If defined to be C<1>, libev will compile in support for the Linux inotify	4508	If defined to be C<1>, libev will compile in support for the Linux inotify
4105	interface to speed up C<ev_stat> watchers. Its actual availability will	4509	interface to speed up C<ev_stat> watchers. Its actual availability will
4106	be detected at runtime. If undefined, it will be enabled if the headers	4510	be detected at runtime. If undefined, it will be enabled if the headers
4107	indicate GNU/Linux + Glibc 2.4 or newer, otherwise disabled.	4511	indicate GNU/Linux + Glibc 2.4 or newer, otherwise disabled.
4108		4512
		4513	=item EV_NO_SMP
		4514
		4515	If defined to be C<1>, libev will assume that memory is always coherent
		4516	between threads, that is, threads can be used, but threads never run on
		4517	different cpus (or different cpu cores). This reduces dependencies
		4518	and makes libev faster.
		4519
		4520	=item EV_NO_THREADS
		4521
		4522	If defined to be C<1>, libev will assume that it will never be called
		4523	from different threads, which is a stronger assumption than C<EV_NO_SMP>,
		4524	above. This reduces dependencies and makes libev faster.
		4525
4109	=item EV_ATOMIC_T	4526	=item EV_ATOMIC_T
4110		4527
4111	Libev requires an integer type (suitable for storing C<0> or C<1>) whose	4528	Libev requires an integer type (suitable for storing C<0> or C<1>) whose
4112	access is atomic with respect to other threads or signal contexts. No such	4529	access is atomic and serialised with respect to other threads or signal
4113	type is easily found in the C language, so you can provide your own type	4530	contexts. No such type is easily found in the C language, so you can
4114	that you know is safe for your purposes. It is used both for signal handler "locking"	4531	provide your own type that you know is safe for your purposes. It is used
4115	as well as for signal and thread safety in C<ev_async> watchers.	4532	both for signal handler "locking" as well as for signal and thread safety
		4533	in C<ev_async> watchers.
4116		4534
4117	In the absence of this define, libev will use C<sig_atomic_t volatile>	4535	In the absence of this define, libev will use C<sig_atomic_t volatile>
4118	(from F<signal.h>), which is usually good enough on most platforms.	4536	(from F<signal.h>), which is usually good enough on most platforms,
		4537	although strictly speaking using a type that also implies a memory fence
		4538	is required.
4119		4539
4120	=item EV_H (h)	4540	=item EV_H (h)
4121		4541
4122	The name of the F<ev.h> header file used to include it. The default if	4542	The name of the F<ev.h> header file used to include it. The default if
4123	undefined is C<"ev.h"> in F<event.h>, F<ev.c> and F<ev++.h>. This can be	4543	undefined is C<"ev.h"> in F<event.h>, F<ev.c> and F<ev++.h>. This can be
…		…
4147	will have the C<struct ev_loop *> as first argument, and you can create	4567	will have the C<struct ev_loop *> as first argument, and you can create
4148	additional independent event loops. Otherwise there will be no support	4568	additional independent event loops. Otherwise there will be no support
4149	for multiple event loops and there is no first event loop pointer	4569	for multiple event loops and there is no first event loop pointer
4150	argument. Instead, all functions act on the single default loop.	4570	argument. Instead, all functions act on the single default loop.
4151		4571
		4572	Note that C<EV_DEFAULT> and C<EV_DEFAULT_> will no longer provide a
		4573	default loop when multiplicity is switched off - you always have to
		4574	initialise the loop manually in this case.
		4575
4152	=item EV_MINPRI	4576	=item EV_MINPRI
4153		4577
4154	=item EV_MAXPRI	4578	=item EV_MAXPRI
4155		4579
4156	The range of allowed priorities. C<EV_MINPRI> must be smaller or equal to	4580	The range of allowed priorities. C<EV_MINPRI> must be smaller or equal to
…		…
4253		4677
4254	With an intelligent-enough linker (gcc+binutils are intelligent enough	4678	With an intelligent-enough linker (gcc+binutils are intelligent enough
4255	when you use C<-Wl,--gc-sections -ffunction-sections>) functions unused by	4679	when you use C<-Wl,--gc-sections -ffunction-sections>) functions unused by
4256	your program might be left out as well - a binary starting a timer and an	4680	your program might be left out as well - a binary starting a timer and an
4257	I/O watcher then might come out at only 5Kb.	4681	I/O watcher then might come out at only 5Kb.
		4682
		4683	=item EV_API_STATIC
		4684
		4685	If this symbol is defined (by default it is not), then all identifiers
		4686	will have static linkage. This means that libev will not export any
		4687	identifiers, and you cannot link against libev anymore. This can be useful
		4688	when you embed libev, only want to use libev functions in a single file,
		4689	and do not want its identifiers to be visible.
		4690
		4691	To use this, define C<EV_API_STATIC> and include F<ev.c> in the file that
		4692	wants to use libev.
		4693
		4694	This option only works when libev is compiled with a C compiler, as C++
		4695	doesn't support the required declaration syntax.
4258		4696
4259	=item EV_AVOID_STDIO	4697	=item EV_AVOID_STDIO
4260		4698
4261	If this is set to C<1> at compiletime, then libev will avoid using stdio	4699	If this is set to C<1> at compiletime, then libev will avoid using stdio
4262	functions (printf, scanf, perror etc.). This will increase the code size	4700	functions (printf, scanf, perror etc.). This will increase the code size
…		…
4406	And a F<ev_cpp.C> implementation file that contains libev proper and is compiled:	4844	And a F<ev_cpp.C> implementation file that contains libev proper and is compiled:
4407		4845
4408	#include "ev_cpp.h"	4846	#include "ev_cpp.h"
4409	#include "ev.c"	4847	#include "ev.c"
4410		4848
4411	=head1 INTERACTION WITH OTHER PROGRAMS OR LIBRARIES	4849	=head1 INTERACTION WITH OTHER PROGRAMS, LIBRARIES OR THE ENVIRONMENT
4412		4850
4413	=head2 THREADS AND COROUTINES	4851	=head2 THREADS AND COROUTINES
4414		4852
4415	=head3 THREADS	4853	=head3 THREADS
4416		4854
…		…
4467	default loop and triggering an C<ev_async> watcher from the default loop	4905	default loop and triggering an C<ev_async> watcher from the default loop
4468	watcher callback into the event loop interested in the signal.	4906	watcher callback into the event loop interested in the signal.
4469		4907
4470	=back	4908	=back
4471		4909
4472	=head4 THREAD LOCKING EXAMPLE	4910	See also L<THREAD LOCKING EXAMPLE>.
4473
4474	Here is a fictitious example of how to run an event loop in a different
4475	thread than where callbacks are being invoked and watchers are
4476	created/added/removed.
4477
4478	For a real-world example, see the C<EV::Loop::Async> perl module,
4479	which uses exactly this technique (which is suited for many high-level
4480	languages).
4481
4482	The example uses a pthread mutex to protect the loop data, a condition
4483	variable to wait for callback invocations, an async watcher to notify the
4484	event loop thread and an unspecified mechanism to wake up the main thread.
4485
4486	First, you need to associate some data with the event loop:
4487
4488	typedef struct {
4489	mutex_t lock; /* global loop lock */
4490	ev_async async_w;
4491	thread_t tid;
4492	cond_t invoke_cv;
4493	} userdata;
4494
4495	void prepare_loop (EV_P)
4496	{
4497	// for simplicity, we use a static userdata struct.
4498	static userdata u;
4499
4500	ev_async_init (&u->async_w, async_cb);
4501	ev_async_start (EV_A_ &u->async_w);
4502
4503	pthread_mutex_init (&u->lock, 0);
4504	pthread_cond_init (&u->invoke_cv, 0);
4505
4506	// now associate this with the loop
4507	ev_set_userdata (EV_A_ u);
4508	ev_set_invoke_pending_cb (EV_A_ l_invoke);
4509	ev_set_loop_release_cb (EV_A_ l_release, l_acquire);
4510
4511	// then create the thread running ev_loop
4512	pthread_create (&u->tid, 0, l_run, EV_A);
4513	}
4514
4515	The callback for the C<ev_async> watcher does nothing: the watcher is used
4516	solely to wake up the event loop so it takes notice of any new watchers
4517	that might have been added:
4518
4519	static void
4520	async_cb (EV_P_ ev_async *w, int revents)
4521	{
4522	// just used for the side effects
4523	}
4524
4525	The C<l_release> and C<l_acquire> callbacks simply unlock/lock the mutex
4526	protecting the loop data, respectively.
4527
4528	static void
4529	l_release (EV_P)
4530	{
4531	userdata *u = ev_userdata (EV_A);
4532	pthread_mutex_unlock (&u->lock);
4533	}
4534
4535	static void
4536	l_acquire (EV_P)
4537	{
4538	userdata *u = ev_userdata (EV_A);
4539	pthread_mutex_lock (&u->lock);
4540	}
4541
4542	The event loop thread first acquires the mutex, and then jumps straight
4543	into C<ev_run>:
4544
4545	void *
4546	l_run (void *thr_arg)
4547	{
4548	struct ev_loop *loop = (struct ev_loop *)thr_arg;
4549
4550	l_acquire (EV_A);
4551	pthread_setcanceltype (PTHREAD_CANCEL_ASYNCHRONOUS, 0);
4552	ev_run (EV_A_ 0);
4553	l_release (EV_A);
4554
4555	return 0;
4556	}
4557
4558	Instead of invoking all pending watchers, the C<l_invoke> callback will
4559	signal the main thread via some unspecified mechanism (signals? pipe
4560	writes? C<Async::Interrupt>?) and then waits until all pending watchers
4561	have been called (in a while loop because a) spurious wakeups are possible
4562	and b) skipping inter-thread-communication when there are no pending
4563	watchers is very beneficial):
4564
4565	static void
4566	l_invoke (EV_P)
4567	{
4568	userdata *u = ev_userdata (EV_A);
4569
4570	while (ev_pending_count (EV_A))
4571	{
4572	wake_up_other_thread_in_some_magic_or_not_so_magic_way ();
4573	pthread_cond_wait (&u->invoke_cv, &u->lock);
4574	}
4575	}
4576
4577	Now, whenever the main thread gets told to invoke pending watchers, it
4578	will grab the lock, call C<ev_invoke_pending> and then signal the loop
4579	thread to continue:
4580
4581	static void
4582	real_invoke_pending (EV_P)
4583	{
4584	userdata *u = ev_userdata (EV_A);
4585
4586	pthread_mutex_lock (&u->lock);
4587	ev_invoke_pending (EV_A);
4588	pthread_cond_signal (&u->invoke_cv);
4589	pthread_mutex_unlock (&u->lock);
4590	}
4591
4592	Whenever you want to start/stop a watcher or do other modifications to an
4593	event loop, you will now have to lock:
4594
4595	ev_timer timeout_watcher;
4596	userdata *u = ev_userdata (EV_A);
4597
4598	ev_timer_init (&timeout_watcher, timeout_cb, 5.5, 0.);
4599
4600	pthread_mutex_lock (&u->lock);
4601	ev_timer_start (EV_A_ &timeout_watcher);
4602	ev_async_send (EV_A_ &u->async_w);
4603	pthread_mutex_unlock (&u->lock);
4604
4605	Note that sending the C<ev_async> watcher is required because otherwise
4606	an event loop currently blocking in the kernel will have no knowledge
4607	about the newly added timer. By waking up the loop it will pick up any new
4608	watchers in the next event loop iteration.
4609		4911
4610	=head3 COROUTINES	4912	=head3 COROUTINES
4611		4913
4612	Libev is very accommodating to coroutines ("cooperative threads"):	4914	Libev is very accommodating to coroutines ("cooperative threads"):
4613	libev fully supports nesting calls to its functions from different	4915	libev fully supports nesting calls to its functions from different
…		…
4778	requires, and its I/O model is fundamentally incompatible with the POSIX	5080	requires, and its I/O model is fundamentally incompatible with the POSIX
4779	model. Libev still offers limited functionality on this platform in	5081	model. Libev still offers limited functionality on this platform in
4780	the form of the C<EVBACKEND_SELECT> backend, and only supports socket	5082	the form of the C<EVBACKEND_SELECT> backend, and only supports socket
4781	descriptors. This only applies when using Win32 natively, not when using	5083	descriptors. This only applies when using Win32 natively, not when using
4782	e.g. cygwin. Actually, it only applies to the microsofts own compilers,	5084	e.g. cygwin. Actually, it only applies to the microsofts own compilers,
4783	as every compielr comes with a slightly differently broken/incompatible	5085	as every compiler comes with a slightly differently broken/incompatible
4784	environment.	5086	environment.
4785		5087
4786	Lifting these limitations would basically require the full	5088	Lifting these limitations would basically require the full
4787	re-implementation of the I/O system. If you are into this kind of thing,	5089	re-implementation of the I/O system. If you are into this kind of thing,
4788	then note that glib does exactly that for you in a very portable way (note	5090	then note that glib does exactly that for you in a very portable way (note
…		…
4921		5223
4922	The type C<double> is used to represent timestamps. It is required to	5224	The type C<double> is used to represent timestamps. It is required to
4923	have at least 51 bits of mantissa (and 9 bits of exponent), which is	5225	have at least 51 bits of mantissa (and 9 bits of exponent), which is
4924	good enough for at least into the year 4000 with millisecond accuracy	5226	good enough for at least into the year 4000 with millisecond accuracy
4925	(the design goal for libev). This requirement is overfulfilled by	5227	(the design goal for libev). This requirement is overfulfilled by
4926	implementations using IEEE 754, which is basically all existing ones. With	5228	implementations using IEEE 754, which is basically all existing ones.
		5229
4927	IEEE 754 doubles, you get microsecond accuracy until at least 2200.	5230	With IEEE 754 doubles, you get microsecond accuracy until at least the
		5231	year 2255 (and millisecond accuracy till the year 287396 - by then, libev
		5232	is either obsolete or somebody patched it to use C<long double> or
		5233	something like that, just kidding).
4928		5234
4929	=back	5235	=back
4930		5236
4931	If you know of other additional requirements drop me a note.	5237	If you know of other additional requirements drop me a note.
4932		5238
…		…
4994	=item Processing ev_async_send: O(number_of_async_watchers)	5300	=item Processing ev_async_send: O(number_of_async_watchers)
4995		5301
4996	=item Processing signals: O(max_signal_number)	5302	=item Processing signals: O(max_signal_number)
4997		5303
4998	Sending involves a system call I<iff> there were no other C<ev_async_send>	5304	Sending involves a system call I<iff> there were no other C<ev_async_send>
4999	calls in the current loop iteration. Checking for async and signal events	5305	calls in the current loop iteration and the loop is currently
		5306	blocked. Checking for async and signal events involves iterating over all
5000	involves iterating over all running async watchers or all signal numbers.	5307	running async watchers or all signal numbers.
5001		5308
5002	=back	5309	=back
5003		5310
5004		5311
5005	=head1 PORTING FROM LIBEV 3.X TO 4.X	5312	=head1 PORTING FROM LIBEV 3.X TO 4.X
…		…
5122	The physical time that is observed. It is apparently strictly monotonic :)	5429	The physical time that is observed. It is apparently strictly monotonic :)
5123		5430
5124	=item wall-clock time	5431	=item wall-clock time
5125		5432
5126	The time and date as shown on clocks. Unlike real time, it can actually	5433	The time and date as shown on clocks. Unlike real time, it can actually
5127	be wrong and jump forwards and backwards, e.g. when the you adjust your	5434	be wrong and jump forwards and backwards, e.g. when you adjust your
5128	clock.	5435	clock.
5129		5436
5130	=item watcher	5437	=item watcher
5131		5438
5132	A data structure that describes interest in certain events. Watchers need	5439	A data structure that describes interest in certain events. Watchers need
…		…
5135	=back	5442	=back
5136		5443
5137	=head1 AUTHOR	5444	=head1 AUTHOR
5138		5445
5139	Marc Lehmann <libev@schmorp.de>, with repeated corrections by Mikael	5446	Marc Lehmann <libev@schmorp.de>, with repeated corrections by Mikael
5140	Magnusson and Emanuele Giaquinta.	5447	Magnusson and Emanuele Giaquinta, and minor corrections by many others.
5141		5448

Diff Legend

–	Removed lines
+	Added lines
<	Changed lines
>	Changed lines