ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/EV/Makefile.PL
(Generate patch)

Comparing EV/Makefile.PL (file contents):
Revision 1.63 by root, Fri Jul 12 06:18:37 2019 UTC vs.
Revision 1.64 by root, Tue Nov 19 13:08:54 2019 UTC

7 scalar grep -r "$_/$_[0]", $Config{usrinc}, split / /, $Config{incpth} 7 scalar grep -r "$_/$_[0]", $Config{usrinc}, split / /, $Config{incpth}
8} 8}
9 9
10my $DEFINE; 10my $DEFINE;
11 11
12unless (-e "libev/ev_linuxaio.c") { 12unless (-e "libev/ev_iouring.c") {
13 print <<EOF; 13 print <<EOF;
14 14
15*** 15***
16*** ERROR: libev is missing or damaged. If you used a CVS check-out of EV, 16*** ERROR: libev is missing or damaged. If you used a CVS check-out of EV,
17*** you also have to check-out the "libev" module from the same CVS 17*** you also have to check-out the "libev" module from the same CVS
185print <<EOF; 185print <<EOF;
186 186
187*** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** 187*** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***
188 188
189 189
190Similarly to the epoll backend above, EV can take advantage of kqueue on 190Linux 4.19 introduced another event polling interface, "io_uring". While
191many BSD systems. Support for kqueue will be detected at runtime, with a 191this API is far superior to epoll and almost rivals linuxaio, it also
192safe fallback to other methods when it cannot be used. 192suffers from the same issues as kqueue typically does: only a subset of
193file types are supported (as of 5.2). It is also very buggy still, and
194most importantly, very very slow for most workloads. Therefore, this
195backend is not used by default, even when it is compiled in, and you have
196to request it explicitly, e.g. with LIBEV_FLAGS=128. If unsure, accept the
197default.
193 198
199EOF
200
201my $can_iouring = have_inc "linux/fs.h";
202$can_iouring = $ENV{EV_IOURING} if exists $ENV{EV_IOURING};
203$can_iouring = 0 + (prompt ("Enable linux io_uring backend (y/n)?", $can_iouring ? "y" : "n") =~ /[yY]/);
204$DEFINE .= " -DEV_USE_IOURING=$can_iouring";
205
206if ($can_iouring) {
207print <<EOF;
208
209*** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***
210
211
212The previously mentioned Linux io_uring is experimental and will not be
213used unless requested explicitly. You can, howeer, choose to make ti a
214recommended basckend, which means it will be chosen if available even when
215not explicitly asked for, in preference to epoll on GNU/Linux. This option
216is likely temporary. When unsure, accept the default.
217
218EOF
219
220my $recommend_iouring = 0;
221$recommend_iouring = $ENV{EV_RECOMMEND_IOURING} if exists $ENV{EV_RECOMMEND_IOURING};
222$recommend_iouring = 0 + (prompt ("Treat io_uring as a recommended backend (y/n)?", $recommend_iouring ? "y" : "n") =~ /[yY]/);
223$DEFINE .= " -DEV_RECOMMEND_IOURING=$recommend_iouring";
224}
225
226print <<EOF;
227
228*** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***
229
230
231EV can take advantage of kqueue on many BSD systems. Support for kqueue
232will be detected at runtime, with a safe fallback to other methods when it
233cannot be used.
234
194Note that kqueue is broken on most operating systems, so by default it 235Note that kqueue is subtly broken on most operating systems, so by default
195won't be used on many platforms, but you can still create your own event 236it won't be used on many platforms, but you can still create your own
196loop with kqueue backend if you ask specifically for it. 237event loop with kqueue backend if you ask specifically for it.
197 238
198Here is what we know: 239Here is what we know:
199 240
200NetBSD: partially working in at least 3.1 and later. Yeah! :) 241NetBSD: partially working in at least 3.1 and later. Yeah! :)
201FreeBSD: broken on at least 6.2-STABLE, spotty in later versions, 242FreeBSD: broken on at least 6.2-STABLE, spotty in later versions,

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines