_kqsub_free assumes the caller has called _kqsub_cancel before calling
it. It checks if both 'deps' and 'fd' have been freed and aborts when
the condition is not met. Since the only caller of _kqsub_free is
g_kqueue_file_monitor_finalize, which does call _kqsub_cancel before
calling _kqsub_free, it seems to be correct for _kqsub_free to assert
values of these two members there.
However, it is possible for _kqsub_cancel to return early without
freeing any resource _kqsub_free expects to be freed. When the kevent
call fails, _kqsub_cancel does not free anything and _kqsub_free aborts
with assertion failure. This is an unexpected behavior, and it can be
fixed by always freeing resources in _kqsub_cancel.
Fixes: https://gitlab.gnome.org/GNOME/glib/issues/1935
Previously, kqueue file monitor only add event sources for directories
regardless of the type of the file being monitored. Doing so may be
possible on inotify, but it is not sufficient on kqueue. Watching a
directory on kqueue doesn't report changes made to files under it, and
we must watch files themselves to get notified. This problem is fixed
by adding a second watch for non-directory file monitors, and the result
is that we are now able to receive 'CHANGED' and 'ATTRIBUTE_CHANGED'
events for non-directory files.
Since having two watches on one file monitor requires many code changes
to work properly, this commit also changes the following things:
- NOTE_ALL macro is now replaced by note_all inline function. Since the
kqueue backend is shared by all BSD operating systems, there are a
few difference between these systems. It is easier to do '#ifdef'
check in a function than in a macro.
- Both g_kqueue_file_monitor_callback and g_kqueue_file_monitor_cancel
now holds a lock before accessing kqueue_sub structs. This fixes a
crash when these two functions are called from different threads,
causing g_kqueue_file_monitor_callback to access freed memory.
- 'mask' variable in g_kqueue_file_monitor_callback is now removed.
The usage of 'mask' was wrong because of the 'mask > 0' check.
'CHANGED' event has value 0 so the 'mask > 0' check made it
impossible to emit 'CHANGED' events.
- kqueue-missing scans can now be triggered from the kqueue event
callback instead of always waiting for 4 seconds.
- Don't remove a file from kqueue on unlink unless its hard link count
has dropped to zero.
- Don't use 'else if' in the check of 'fflags'. It is possible for a
kevent to have multiple flags set.
- Don't use g_file_monitor_emit_event directly. Always use
g_file_monitor_source_handle_event to report events.
Events submitted to g_file_monitor_emit_event are delivered
immediately, but events sent to g_file_monitor_source_handle_event
are scheduled by GLocalFileMonitor. If we mix the two, the order of
events will be wrong and tests will fail.
- Report 'CHANGES_DONE_HINT' immediately after 'CREATED' if the file
created is not a regular file. This is copied from ih_event_callback.
The check in _ke_is_excluded, which causes GKqueueFileMonitor to
fallback to GPollFileMonitor when it returns TRUE, was made to prevent
file monitor from blocking unmount of removable drives on systems not
supporting O_EVTONLY flag in open. However, since g_mount_can_unmount
always returns TRUE on Unix-like platforms, the check always returns
TRUE on non-standard mount points, which is very likely to cause all
programs on the desktop to use the polling fallback if GNOME is
installed in a different prefix for development. This makes the desktop
sluggish and results in bad developer experience on *BSD.
Removing the event and closing the related file descriptor must be
done first to make sure the kqueue subsystem delete pending events.
The timeout must be disarmed before freeing the directory dependency
list otherwise it might populate it again.
https://bugzilla.gnome.org/show_bug.cgi?id=795193
- Stop using a custom thread for listening to kqueue(2) events. Instead
call kevent(2) in non blocking mode in a monitor callback. Under the
hood poll(2) is used to figure out if new events are available.
- Do not use a socketpair with a custom protocol requiring 2 supplementary
context switches per event to commicate between multiple threads. Calling
kevent(2), in non blocking mode, to add/remove events is fine from any
context.
- Add kqueue(2) events without the EV_ONESHOT flag. This removes a race
where some notifications were lost because events had to be re-added for
every new notification.
- Get rid of the global hash table and its associated lock and races. Use
the 'cookie' argument of kevent(2) to pass the associated descriptor when
registering an event.
- Fix _kh_file_appeared_cb() by properly passing a monitor instead of a
source to g_file_monitor_emit_event().
- Properly refcount sources.
- Remove a lot of abstraction making it harder to fix the remaining issues.
https://bugzilla.gnome.org/show_bug.cgi?id=739424
Cancellation of GPollFileMonitor is now handled correctly (in the sense
that no further signals will follow) but let's be extra paranoid and
disconnect our handler anyway, for good measure.
https://bugzilla.gnome.org/show_bug.cgi?id=739424
This is the bare minimal effort. This seems not to crash immediately,
but it definitely needs some better testing.
The backend is not in good shape. It could use some serious work.
Written by Dmitry Matveev as part of GSoC 2011:
http://netbsd-soc.sourceforge.net/projects/kqueue4gio/
This brings native file monitoring support on systems supporting kqueue(3)
(all BSDs) and remove the need to rely on the unmaintained gamin software.
The backend adds GKqueueDirectoryMonitor and GKqueueFileMonitor.
Some parts rewritten by myself (to prevent needing a configuration file).
Helpful inputs from Colin Walters and Simon McVittie.
https://bugzilla.gnome.org/show_bug.cgi?id=679793
