dbus-glib subscribes to all NameOwnerChanged signals for each DBusGProxy instance (or possibly just the ones bound to a unique name), so all dbus-glib applications are woken up whenever any application joins or leaves the bus. It should subscribe using "arg0" matching, like dbus-python does, so only signals it cares about cause a wakeup. It should also provide API for clients to do this.
watch_name_owner(bus_name, callback, the usual closure stuff): like tp_dbus_daemon_watch_name_owner in telepathy-glib or dbus.Bus.watch_name_owner in dbus-python. Call a user-supplied callback when the name owner has been discovered, then call it again every time the name owner changes. Some way to disconnect from the notification should be provided.
As a small optimization, if the name given was a unique name, we could automatically disconnect the notification when it fell off the bus (since the whole point is that unique names aren't re-used).
As a nasty effect of this issue, application developers can afford to be sloppy with indiscriminately connecting to NameOwnerChanged and point to dbus-glib in justification. So the later this bug gets fixed, the more applications will commit the same mistake and won't be singled out as performance drains. I know at least one example apart from bug #14184.
Created attachment 22614 [details] [review]
Provisional patch - passes the current test suite, but I haven't double checked the wakeups yet.
On the subject of a watch_name_owner API - this would be very nice, I agree; but right now in a lot of cases it is sufficient to use dbus_g_proxy_new_for_name_owner and connect to the "destroy" signal.
I've verified this patch prevents wakeups. Test case was attaching strace to bluetooth-applet and then starting up dbus-monitor --session. With patch, every invocation of dbus-monitor woke the applet up out of poll() to read the NameOwnerChanged signal. With the patch, it does not.
Oops, patch writing race condition. My patch is substantially similar, but more complicated in that it avoids adding duplicate match rules, but maybe this doesn't win us anything. Your patch looks fine to me.
Ahh, apologies for the duplication; both of us are probably at fault here for not claiming the bugzilla mutex =)
But thanks for the review! I hadn't considered the duplicate issue; my intuition is that if it was a real-world problem, it'd be simpler to have the bus daemon keep a reference count for matches instead of a plain linked list.
I've pushed this patch to master. I've been running it on my system since posting it, but if anyone else could give it some testing on their own system before the next release I'd appreciate it!
Author: Colin Walters <firstname.lastname@example.org>
Date: Thu Feb 5 11:17:15 2009 -0500
Bug 14183 - Listen to NameOwnerChanged using arg0 matching
This is more efficient - we avoid waking up every dbus-glib using process
when a process joins or leaves the bus.
on Mar 26, 2017 at 01:32:41.
(provided by the Example extension).