fake_inotify + kqnotify: rewrite and combine code

KQNotify is now a subclass of FakeInotify since they're both
faking a subset of inotify; and both require directory scanning
via readdir() to detect new/deleted files.

ctime is no longer used with per-file stat to detect new files
with kevent.  That proved too unreliable either due to low
time resolution of the NetBSD/OpenBSD VFS and/or
Time::HiRes::stat being constrained by floating point to
represent `struct timespec', so instead we fuzz the time a bit
if the ctime is recent and merely compare filenames off readdir.

This fixes t/fake_inotify.t and t/kqnotify.t failures under NetBSD
and also removes workarounds for OpenBSD in t/kqnotify.t.  It
also allows us to to remove delays in tests by being more
aggressive in picking up new/deleted files in watch directories
by adjusting the time to scan if the ctime is recent.

This ought to may improve real-world reliability on all *BSDs
regardless of whether IO::KQueue is installed.

1 files changed, 1 insertions, 4 deletions

diff --git a/t/dir_idle.t b/t/dir_idle.t
index 50e1dd27..9e66cc58 100644
--- a/t/dir_idle.t
+++ b/t/dir_idle.t
@@ -14,14 +14,12 @@ $di->add_watches(["$tmpdir/a", "$tmpdir/c"], 1);
 PublicInbox::DS->SetLoopTimeout(1000);
 my $end = 3 + now;
 local @PublicInbox::DS::post_loop_do = (sub { scalar(@x) == 0 && now < $end });
-tick(0.011);
 rmdir("$tmpdir/a/b") or xbail "rmdir $!";
 PublicInbox::DS::event_loop();
-is(scalar(@x), 1, 'got an event') and
+is(scalar(@x), 1, 'got an rmdir event') and
         is($x[0]->[0]->fullname, "$tmpdir/a/b", 'got expected fullname') and
         ok($x[0]->[0]->IN_DELETE, 'IN_DELETE set');
 
-tick(0.011);
 rmdir("$tmpdir/a") or xbail "rmdir $!";
 @x = ();
 $end = 3 + now;
@@ -30,7 +28,6 @@ is(scalar(@x), 1, 'got an event') and
         is($x[0]->[0]->fullname, "$tmpdir/a", 'got expected fullname') and
         ok($x[0]->[0]->IN_DELETE_SELF, 'IN_DELETE_SELF set');
 
-tick(0.011);
 rename("$tmpdir/c", "$tmpdir/j") or xbail "rmdir $!";
 @x = ();
 $end = 3 + now;