500 lines
26 KiB

Revision history for libev, a high-performance and full-featured event loop.
TODO: ev_loop_wakeup
TODO: EV_STANDALONE == NO_HASSEL (do not use clock_gettime in ev_standalone)
TODO: faq, process a thing in each iteration
TODO: dbeugging tips, ev_verify, ev_init twice
TODO: ev_break for immediate exit (EVBREAK_NOW?)
TODO: ev_feed_child_event
TODO: document the special problem of signals around fork.
TODO: store pid for each signal
TODO: document file descriptor usage per loop
TODO: store loop pid_t and compare isndie signal handler,store 1 for same, 2 for differign pid, clean up in loop_fork
TODO: embed watchers need updating when fd changes
TODO: document portability requirements for atomic pointer access
TODO: possible cb aliasing?
TODO: document requirements for function pointers and calling conventions.
- ev.h wasn't valid C++ anymore, which tripped compilers other than
clang, msvc or gcc (analyzed by Raphael 'kena' Poss).
- when not using autoconf, tighten the check for clock_gettime and related
4.18 Fri Sep 5 17:55:26 CEST 2014
- events on files were not always generated properly with the
epoll backend (testcase by Assaf Inbal).
- mark event pipe fd as cloexec after a fork (analyzed by Sami Farin).
- (ecb) support m68k, m88k and sh (patch by Miod Vallat).
- use a reasonable fallback for EV_NSIG instead of erroring out
when we can't detect the signal set size.
- in the absence of autoconf, do not use the clock syscall
on glibc >= 2.17 (avoids the syscall AND -lrt on systems
doing clock_gettime in userspace).
- ensure extern "C" function pointers are used for externally-visible
loop callbacks (not watcher callbacks yet).
- (ecb) work around memory barriers and volatile apparently both being
broken in visual studio 2008 and later (analysed and patch by Nicolas Noble).
4.15 Fri Mar 1 12:04:50 CET 2013
- destroying a non-default loop would stop the global waitpid
watcher (Denis Bilenko).
- queueing pending watchers of higher priority from a watcher now invokes
them in a timely fashion (reported by Denis Bilenko).
- add throw() to all libev functions that cannot throw exceptions, for
further code size decrease when compiling for C++.
- add throw () to callbacks that must not throw exceptions (allocator,
syserr, loop acquire/release, periodic reschedule cbs).
- fix event_base_loop return code, add event_get_callback, event_base_new,
event_base_get_method calls to improve libevent 1.x emulation and add
some libevent 2.x functionality (based on a patch by Jeff Davey).
- add more memory fences to fix a bug reported by Jeff Davey. Better
be overfenced than underprotected.
- ev_run now returns a boolean status (true meaning watchers are
still active).
- ev_once: undef EV_ERROR in ev_kqueue.c, to avoid clashing with
libev's EV_ERROR (reported by 191919).
- (ecb) add memory fence support for xlC (Darin McBride).
- (ecb) add memory fence support for gcc-mips (Anton Kirilov).
- (ecb) add memory fence support for gcc-alpha (Christian Weisgerber).
- work around some kernels losing file descriptors by leaking
the kqueue descriptor in the child.
- work around linux inotify not reporting IN_ATTRIB changes for directories
in many cases.
- include sys/syscall.h instead of plain syscall.h.
- check for io watcher loops in ev_verify, check for the most
common reported usage bug in ev_io_start.
- choose socket vs. WSASocket at compiletime using EV_USE_WSASOCKET.
- always use WSASend/WSARecv directly on windows, hoping that this
works in all cases (unlike read/write/send/recv...).
- try to detect signals around a fork faster (test program by
Denis Bilenko).
- work around recent glibc versions that leak memory in realloc.
- rename ev::embed::set to ev::embed::set_embed to avoid clashing
the watcher base set (loop) method.
- rewrite the async/signal pipe logic to always keep a valid fd, which
simplifies (and hopefully correctifies :) the race checking
on fork, at the cost of one extra fd.
- add fat, msdos, jffs2, ramfs, ntfs and btrfs to the list of
inotify-supporting filesystems.
- move orig_CFLAGS assignment to after AC_INIT, as newer autoconf
versions ignore it before
- add some untested android support.
- enum expressions must be of type int (reported by Juan Pablo L).
4.11 Sat Feb 4 19:52:39 CET 2012
- INCOMPATIBLE CHANGE: ev_timer_again now clears the pending status, as
was documented already, but not implemented in the repeating case.
- new compiletime symbols: EV_NO_SMP and EV_NO_THREADS.
- fix a race where the workaround against the epoll fork bugs
caused signals to not be handled anymore.
- correct backend_fudge for most backends, and implement a windows
specific workaround to avoid looping because we call both
select and Sleep, both with different time resolutions.
- document range and guarantees of ev_sleep.
- document reasonable ranges for periodics interval and offset.
- rename backend_fudge to backend_mintime to avoid future confusion :)
- change the default periodic reschedule function to hopefully be more
exact and correct even in corner cases or in the far future.
- do not rely on -lm anymore: use it when available but use our
own floor () if it is missing. This should make it easier to embed,
as no external libraries are required.
- strategically import macros from libecb and mark rarely-used functions
as cache-cold (saving almost 2k code size on typical amd64 setups).
- add Symbols.ev and Symbols.event files, that were missing.
- fix backend_mintime value for epoll (was 1/1024, is 1/1000 now).
- fix #3 "be smart about timeouts" to not "deadlock" when
timeout == now, also improve the section overall.
- support new EV_API_STATIC mode to make all libev symbols
- supply default CFLAGS of -g -O3 with gcc when original CFLAGS
were empty.
4.04 Wed Feb 16 09:01:51 CET 2011
- fix two problems in the native win32 backend, where reuse of fd's
with different underlying handles caused handles not to be removed
or added to the select set (analyzed and tested by Bert Belder).
- do no rely on ceil() in ev_e?poll.c.
- backport libev to HP-UX versions before 11 v3.
- configure did not detect nanosleep and clock_gettime properly when
they are available in the libc (as opposed to -lrt).
4.03 Tue Jan 11 14:37:25 CET 2011
- officially support polling files with all backends.
- support files, /dev/zero etc. the same way as select in the epoll
backend, by generating events on our own.
- ports backend: work around solaris bug 6874410 and many related ones
(EINTR, maybe more), with no performance loss (note that the solaris
bug report is actually wrong, reality is far more bizarre and broken
than that).
- define EV_READ/EV_WRITE as macros in event.h, as some programs use
#ifdef to test for them.
- new (experimental) function: ev_feed_signal.
- new (to become default) EVFLAG_NOSIGMASK flag.
- new EVBACKEND_MASK symbol.
4.01 Fri Nov 5 21:51:29 CET 2010
- automake fucked it up, apparently, --add-missing -f is not quite enough
to make it update its files, so 4.00 didn't install ev++.h and
event.h on make install. grrr.
- ev_loop(count|depth) didn't return anything (Robin Haberkorn).
- change EV_UNDEF to 0xffffffff to silence some overzealous compilers.
- use "(libev) " prefix for all libev error messages now.
4.00 Mon Oct 25 12:32:12 CEST 2010
- "PORTING FROM LIBEV 3.X TO 4.X" (in ev.pod) is recommended reading.
- ev_embed_stop did not correctly stop the watcher (very good
testcase by Vladimir Timofeev).
- ev_run will now always update the current loop time - it erroneously
didn't when idle watchers were active, causing timers not to fire.
- fix a bug where a timeout of zero caused the timer not to fire
in the libevent emulation (testcase by Péter Szabó).
- applied win32 fixes by Michael Lenaghan (also James Mansion).
- prefer EPOLL_CTL_ADD over EPOLL_CTL_MOD in some more cases, as it
seems the former is *much* faster than the latter.
- linux kernel version detection (for inotify bug workarounds)
did not work properly.
- reduce the number of spurious wake-ups with the ports backend.
- remove dependency on sys/queue.h on freebsd (patch by Vanilla Hsu).
- do async init within ev_async_start, not ev_async_set, which avoids
an API quirk where the set function must be called in the C++ API
even when there is nothing to set.
- add (undocumented) EV_ENABLE when adding events with kqueue,
this might help with OS X, which seems to need it despite documenting
not to need it (helpfully pointed out by Tilghman Lesher).
- do not use poll by default on freebsd, it's broken (what isn't
on freebsd...).
- allow to embed epoll on kernels >= 2.6.32.
- configure now prepends -O3, not appends it, so one can still
override it.
- ev.pod: greatly expanded the portability section, added a porting
section, a description of watcher states and made lots of minor fixes.
- disable poll backend on AIX, the poll header spams the namespace
and it's not worth working around dead platforms (reported
and analyzed by Aivars Kalvans).
- improve header file compatibility of the standalone eventfd code
in an obscure case.
- implement EV_AVOID_STDIO option.
- do not use sscanf to parse linux version number (smaller, faster,
no sscanf dependency).
- new EV_CHILD_ENABLE and EV_SIGNAL_ENABLE configurable settings.
- update libev.m4 HAVE_CLOCK_SYSCALL test for newer glibcs.
- add section on accept() problems to the manpage.
- rename EV_TIMEOUT to EV_TIMER.
- rename ev_loop_count/depth/verify/loop/unloop.
- remove ev_default_destroy and ev_default_fork.
- switch to two-digit minor version.
- work around an apparent gentoo compiler bug.
- define _DARWIN_UNLIMITED_SELECT. just so.
- use enum instead of #define for most constants.
- improve compatibility to older C++ compilers.
- (experimental) ev_run/ev_default_loop/ev_break/ev_loop_new have now
default arguments when compiled as C++.
- enable automake dependency tracking.
- ev_loop_new no longer leaks memory when loop creation failed.
- new ev_cleanup watcher type.
3.9 Thu Dec 31 07:59:59 CET 2009
- signalfd is no longer used by default and has to be requested
explicitly - this means that easy to catch bugs become hard to
catch race conditions, but the users have spoken.
- point out the unspecified signal mask in the documentation, and
that this is a race condition regardless of EV_SIGNALFD.
- backport inotify code to C89.
- inotify file descriptors could leak into child processes.
- ev_stat watchers could keep an erroneous extra ref on the loop,
preventing exit when unregistering all watchers (testcases
provided by
- implement EV_WIN32_HANDLE_TO_FD and EV_WIN32_CLOSE_FD configuration
symbols to make it easier for apps to do their own fd management.
- support EV_IDLE_ENABLE being disabled in ev++.h
(patch by Didier Spezia).
- take advantage of inotify_init1, if available, to set cloexec/nonblock
on fd creation, to avoid races.
- the signal handling pipe wasn't always initialised under windows
(analysed by lekma).
- changed minimum glibc requirement from glibc 2.9 to 2.7, for
- add missing string.h include (Denis F. Latypoff).
- only replace ev_stat.prev when we detect an actual difference,
so prev is (almost) always different to attr. this might
have caused the problems with 04_stat.t.
- add ev::timer->remaining () method to C++ API.
3.8 Sun Aug 9 14:30:45 CEST 2009
- incompatible change: do not necessarily reset signal handler
to SIG_DFL when a sighandler is stopped.
- ev_default_destroy did not properly free or zero some members,
potentially causing crashes and memory corruption on repeated
ev_default_destroy/ev_default_loop calls.
- take advantage of signalfd on GNU/Linux systems.
- document that the signal mask might be in an unspecified
state when using libev's signal handling.
- take advantage of some GNU/Linux calls to set cloexec/nonblock
on fd creation, to avoid race conditions.
3.7 Fri Jul 17 16:36:32 CEST 2009
- ev_unloop and ev_loop wrongly used a global variable to exit loops,
instead of using a per-loop variable (bug caught by accident...).
- the ev_set_io_collect_interval interpretation has changed.
- add new functionality: ev_set_userdata, ev_userdata,
ev_set_invoke_pending_cb, ev_set_loop_release_cb,
ev_invoke_pending, ev_pending_count, together with a long example
about thread locking.
- add ev_timer_remaining (as requested by Denis F. Latypoff).
- add ev_loop_depth.
- calling ev_unloop in fork/prepare watchers will no longer poll
for new events.
- Denis F. Latypoff corrected many typos in example code snippets.
- honor autoconf detection of EV_USE_CLOCK_SYSCALL, also double-
check that the syscall number is available before trying to
use it (reported by ry@tinyclouds).
- use GetSystemTimeAsFileTime instead of _timeb on windows, for
slightly higher accuracy.
- properly declare ev_loop_verify and ev_now_update even when
- do not compile in any priority code when EV_MAXPRI == EV_MINPRI.
- support EV_MINIMAL==2 for a reduced API.
- actually 0-initialise struct sigaction when installing signals.
- add section on hibernate and stopped processes to ev_timer docs.
3.6 Tue Apr 28 02:49:30 CEST 2009
- multiple timers becoming ready within an event loop iteration
will be invoked in the "correct" order now.
- do not leave the event loop early just because we have no active
watchers, fixing a problem when embedding a kqueue loop
that has active kernel events but no registered watchers
(reported by blacksand blacksand).
- correctly zero the idx values for arrays, so destroying and
reinitialising the default loop actually works (patch by
Malek Hadj-Ali).
- implement ev_suspend and ev_resume.
- new EV_CUSTOM revents flag for use by applications.
- add documentation section about priorities.
- add a glossary to the documentation.
- extend the ev_fork description slightly.
- optimize a jump out of call_pending.
3.53 Sun Feb 15 02:38:20 CET 2009
- fix a bug in event pipe creation on win32 that would cause a
failed assertion on event loop creation (patch by Malek Hadj-Ali).
- probe for CLOCK_REALTIME support at runtime as well and fall
back to gettimeofday if there is an error, to support older
operating systems with newer header files/libraries.
- prefer gettimeofday over clock_gettime with USE_CLOCK_SYSCALL
(default most everywhere), otherwise not.
3.52 Wed Jan 7 21:43:02 CET 2009
- fix compilation of select backend in fd_set mode when NFDBITS is
missing (to get it to compile on QNX, reported by Rodrigo Campos).
- better select-nfds handling when select backend is in fd_set mode.
- diagnose fd_set overruns when select backend is in fd_set mode.
- due to a thinko, instead of disabling everything but
select on the borked OS X platform, everything but select was
allowed (reported by Emanuele Giaquinta).
- actually verify that local and remote port are matching in
libev's socketpair emulation, which makes denial-of-service
attacks harder (but not impossible - it's windows). Make sure
it even works under vista, which thinks that getpeer/sockname
should return fantasy port numbers.
- include "libev" in all assertion messages for potentially
clearer diagnostics.
- event_get_version (libevent compatibility) returned
a useless string instead of the expected version string
(patch by W.C.A. Wijngaards).
3.51 Wed Dec 24 23:00:11 CET 2008
- fix a bug where an inotify watcher was added twice, causing
freezes on hash collisions (reported and analysed by Graham Leggett).
- new config symbol, EV_USE_CLOCK_SYSCALL, to make libev use
a direct syscall - slower, but no dependency on librt et al.
- assume negative return values != -1 signals success of port_getn
(no known failure reports, but it doesn't hurt).
- fork detection in ev_embed now stops and restarts the watcher
- EXPERIMENTAL: default the method to operator () in ev++.h,
to make it nicer to use functors (requested by Benedek László).
- fixed const object callbacks in ev++.h.
- replaced loop_ref argument of watcher.set (loop) by a direct
ev_loop * in ev++.h, to avoid clashes with functor patch.
- do not try to watch the empty string via inotify.
- inotify watchers could be leaked under certain circumstances.
- OS X 10.5 is actually even more broken than earlier versions,
so fall back to select on that piece of garbage.
- fixed some weirdness in the ev_embed documentation.
3.49 Wed Nov 19 11:26:53 CET 2008
- ev_stat watchers will now use inotify as a mere hint on
kernels <2.6.25, or if the filesystem is not in the
"known to be good" list.
- better mingw32 compatibility (it's not as borked as native win32)
(analysed by Roger Pack).
- include stdio.h in the example program, as too many people are
confused by the weird C language otherwise. I guess the next thing
I get told is that the "..." ellipses in the examples don't compile
with their C compiler.
3.48 Thu Oct 30 09:02:37 CET 2008
- further optimise away the EPOLL_CTL_ADD/MOD combo in the epoll
backend by assuming the kernel event mask hasn't changed if
ADD fails with EEXIST.
- work around spurious event notification bugs in epoll by using
a 32-bit generation counter. recreate kernel state if we receive
spurious notifications or unwanted events. this is very costly,
but I didn't come up with this horrible design.
- use memset to initialise most arrays now and do away with the
init functions.
- expand time-out strategies into a "Be smart about timeouts" section.
- drop the "struct" from all ev_watcher declarations in the
documentation and did other clarifications (yeah, it was a mistake
to have a struct AND a function called ev_loop).
- fix a bug where ev_default would not initialise the default
loop again after it was destroyed with ev_default_destroy.
- rename syserr to ev_syserr to avoid name clashes when embedding,
do similar changes for event.c.
3.45 Tue Oct 21 21:59:26 CEST 2008
- disable inotify usage on linux <2.6.25, as it is broken
(reported by Yoann Vandoorselaere).
- ev_stat erroneously would try to add inotify watchers
even when inotify wasn't available (this should only
have a performance impact).
- ev_once now passes both timeout and io to the callback if both
occur concurrently, instead of giving timeouts precedence.
- disable EV_USE_INOTIFY when sys/inotify.h is too old.
3.44 Mon Sep 29 05:18:39 CEST 2008
- embed watchers now automatically invoke ev_loop_fork on the
embedded loop when the parent loop forks.
- new function: ev_now_update (loop).
- verify_watcher was not marked static.
- improve the "associating..." manpage section.
- documentation tweaks here and there.
3.43 Sun Jul 6 05:34:41 CEST 2008
- include more include files on windows to get struct _stati64
(reported by Chris Hulbert, but doesn't quite fix his issue).
- add missing #include <io.h> in ev.c on windows (reported by
Matt Tolton).
3.42 Tue Jun 17 12:12:07 CEST 2008
- work around yet another windows bug: FD_SET actually adds fd's
multiple times to the fd_*SET*, despite official MSN docs claiming
otherwise. Reported and well-analysed by Matt Tolton.
- define NFDBITS to 0 when EV_SELECT_IS_WINSOCKET to make it compile
(reported any analysed by Chris Hulbert).
- fix a bug in ev_ebadf (this function is only used to catch
programming errors in the libev user). reported by Matt Tolton.
- fix a bug in fd_intern on win32 (could lead to compile errors
under some circumstances, but would work correctly if it compiles).
reported by Matt Tolton.
- (try to) work around missing lstat on windows.
- pass in the write fd set as except fd set under windows. windows
is so uncontrollably lame that it requires this. this means that
switching off oobinline is not supported (but tcp/ip doesn't
have oob, so that would be stupid anyways.
- use posix module symbol to auto-detect monotonic clock presence
and some other default values.
3.41 Fri May 23 18:42:54 CEST 2008
- work around an obscure bug in winsocket select: if you
provide only empty fd sets then select returns WSAEINVAL. how sucky.
- improve timer scheduling stability and reduce use of time_epsilon.
- use 1-based 2-heap for EV_MINIMAL, simplifies code, reduces
codesize and makes for better cache-efficiency.
- use 3-based 4-heap for !EV_MINIMAL. this makes better use
of cpu cache lines and gives better growth behaviour than
2-based heaps.
- cache timestamp within heap for !EV_MINIMAL, to avoid random
memory accesses.
- document/add EV_USE_4HEAP and EV_HEAP_CACHE_AT.
- fix a potential aliasing issue in ev_timer_again.
- add/document ev_periodic_at, retract direct access to ->at.
- improve ev_stat docs.
- add portability requirements section.
- fix manpage headers etc.
- normalise WSA error codes to lower range on windows.
- add consistency check code that can be called automatically
or on demand to check for internal structures (ev_loop_verify).
3.31 Wed Apr 16 20:45:04 CEST 2008
- added last minute fix for ev_poll.c by Brandon Black.
3.3 Wed Apr 16 19:04:10 CEST 2008
- event_base_loopexit should return 0 on success
(W.C.A. Wijngaards).
- added linux eventfd support.
- try to autodetect epoll and inotify support
by libc header version if not using autoconf.
- new symbols: EV_DEFAULT_UC and EV_DEFAULT_UC_.
- declare functions defined in ev.h as inline if
C99 or gcc are available.
- enable inlining with gcc versions 2 and 3.
- work around broken poll implementations potentially
not clearing revents field in ev_poll (Brandon Black)
(no such systems are known at this time).
- work around a bug in realloc on openbsd and darwin,
also makes the erroneous valgrind complaints
go away (noted by various people).
- fix ev_async_pending, add c++ wrapper for ev_async
(based on patch sent by Johannes Deisenhofer).
- add sensible set method to ev::embed.
- made integer constants type int in ev.h.
3.2 Wed Apr 2 17:11:19 CEST 2008
- fix a 64 bit overflow issue in the select backend,
by using fd_mask instead of int for the mask.
- rename internal sighandler to avoid clash with very old perls.
- entering ev_loop will not clear the ONESHOT or NONBLOCKING
flags of any outer loops anymore.
- add ev_async_pending.
3.1 Thu Mar 13 13:45:22 CET 2008
- implement ev_async watchers.
- only initialise signal pipe on demand.
- make use of sig_atomic_t configurable.
- improved documentation.
3.0 Mon Jan 28 13:14:47 CET 2008
- API/ABI bump to version 3.0.
- ev++.h includes "ev.h" by default now, not <ev.h>.
- slightly improved documentation.
- speed up signal detection after a fork.
- only optionally return trace status changed in ev_child
- experimental (and undocumented) loop wrappers for ev++.h.
2.01 Tue Dec 25 08:04:41 CET 2007
- separate Changes file.
- fix ev_path_set => ev_stat_set typo.
- remove event_compat.h from the libev tarball.
- change how include files are found.
- doc updates.
- update licenses, explicitly allow for GPL relicensing.
2.0 Sat Dec 22 17:47:03 CET 2007
- new ev_sleep, ev_set_(io|timeout)_collect_interval.
- removed epoll from embeddable fd set.
- fix embed watchers.
- renamed ev_embed.loop to other.
- added exported Symbol tables.
- undefine member wrapper macros at the end of ev.c.
- respect EV_H in ev++.h.
1.86 Tue Dec 18 02:36:57 CET 2007
- fix memleak on loop destroy (not relevant for perl).
1.85 Fri Dec 14 20:32:40 CET 2007
- fix some aliasing issues w.r.t. timers and periodics
(not relevant for perl).
(for historic versions refer to EV/Changes, found in the Perl interface)
0.1 Wed Oct 31 21:31:48 CET 2007
- original version; hacked together in <24h.