reopening the session management discussion...

[Brian J. Tarricone]

Benedikt Meurer wrote:
> Brian J. Tarricone wrote:
>>>>>> * App can never "accidentally" fall out of the session due to a
>>>>>> crash (I guess I'm the biggest offender here). App will always be
>>>>>> started when the DE starts unless the user disables it.
 >>>>>
>>>>> Inmplementing SmRestartImmediately in xfce4-session would solve the
>>>>> problem without loosing the flexibility of session management.
 >>>>
>>>> Well, who wants to do that?  Benny doesn't seem to be around, and I
>>>> feel like there must have been a reason why he didn't implement it
>>>> in the first place.
 >>>
>>> I reallty see no reason other than no time or interest.
 >>
>> I looked through the xfce4-session source for maybe a code comment
>> explaining the omission, but I didn't see anything.  I also thought
>> Benny had talked about it on the ML at some point, but a search of the
>> archives didn't yield anything (perhaps I just suck).
> 
> The SmRestart thing is really tricky to get right and not very useful in
> general. There are several special cases to consider: I.e. stop
> respawning after a number of failed retries (app dumps core on startup),

Well that doesn't sound too hard -- checking WIFSIGNALED() and 
WEXITSTATUS() will tell you how the app died, and you can make an 
intelligent decision based on that.  Hell, even init has code to not 
respawn apps if they quit X number of times in Y seconds, so there's 
certainly precedent for this sort of thing.

> user killing xfwm4 to replace it with metacity (xfce4-session will
> restart xfwm4 immediately), and so on.

Olivier mentioned there's a way around this by making one signal 
'special' (SIGTERM) to the app getting killed such that it can catch 
this signal, reset its SmRestartTypeHint to SmRestartIfRunning (or 
SmRestartNever), and then quit.  I believe xfwm4 relies on this behavior 
in other WMs when passed the --replace option.

> Afterall the XSMP is not very
> well designed for todays needs (gnome-session is probably the most
> complete implementation and it is a total mess for this very reason).
> SmRestartImmediately makes sense for core desktop services only, and for
> those it would be better to have some kind of autostart using a special
> D-Bus service name (allowing replacement). I.e.
> org.freedesktop.WindowManager for the window manager; then i.e. starting
> metacity would simply replace xfwm4 as window manager, simple and
> straight forward. The session manager would simply ask the well-known
> core services (maybe extracting their defined names from some config
> file) for its startup command on logout (IIRC this is already possible
> using the dbus-daemon). For the automatic restart it would be sufficient
> to watch the name registration signals from the bus daemon and Ping the
> service name once a name is lost. It would even be possible to migrate
> existing applications easily providing a simple wrapper binary, which
> does the D-Bus part.

This sounds good to me in theory, but I feel like adding 
SmRestartImmediately support to xfce4-session, and then making xfwm4, 
xfdesktop, xfce4-panel, etc. use it would be doable in a couple hours, 
whereas implementing all this new stuff would take quite a bit more time.

It also suffers some of the same problems you mention with XSMP -- with 
an auto-restart using a dbus Ping, you still have to make sure to not 
respawn the app too many times in a row if it's just core-dumping on 
startup (and here it's just as bad, because you have to wait for it to 
register itself with the session bus, and you can't even easily get its 
PID to tell how it's exiting).  You also have to add some sort of 
session manager API so the app can quit and say "please don't restart 
me; I really want to quit," because we can't yet expect ALL WMs to 
support this, so starting metacity wouldn't register 
org.freedesktop.WindowManager, and then the watchdog process would try 
to ping the bus name again and would restart xfwm4.  And you also have 
to set it up so, when a *new* WM starts, it can tell the dbus-based 
session manager its restart command so it can write a new dbus 
auto-activation file in ~/.local/.

So basically you're suggesting we fix the current problem by 
implementing a new session management protocol that hasn't been designed 
yet and probably implements a good portion of what XSMP implements. 
That's not a small job, by any means.  Not saying I don't agree with 
you; I'd love to see this done, but... who has time right now?

There's also the issue Olivier brought up about how he wants xfwm4 to 
behave correctly outside Xfce as well, so it has to keep registering 
itself with the session (at least until/unless everyone else implements 
our dbus-based session management protocol)...  so we might as well just 
make session management work the best we can without turning 
xfce4-session into a monstrosity.

So, looking at the path of least resistance here...  if I work on a 
patch for xfce4-session to implement SmRestartImmediately support, will 
you accept it?

	-brian