Emacs daemon doesn't terminate on SIGTERM any more

Emacs daemon doesn't terminate on SIGTERM any more

[Ulrich Mueller]

I start Emacs from BZR trunk in daemon mode and connect to it:

   $ emacs -Q --daemon
   ("emacs")
   Starting Emacs daemon.
   $ emacsclient -t

Then I try to kill it from another terminal window:

   $ kill <pid of emacs process>

With Emacs 23.2 the process would properly terminate now. The BZR
version asks for interactive input in the open terminal frame instead:

    The current server still has clients; delete them? (yes or no)

Therefore the process doesn't terminate. Is such a behaviour intended?
IMHO, a process should terminate on reception of a SIGTERM.

Ulrich

Re: Emacs daemon doesn't terminate on SIGTERM any more

[Ulrich Mueller]

>>>>> On Fri, 22 Oct 2010, Ulrich Mueller wrote:

> Then I try to kill it from another terminal window:

>    $ kill <pid of emacs process>

> With Emacs 23.2 the process would properly terminate now. The BZR
> version asks for interactive input in the open terminal frame instead:

>     The current server still has clients; delete them? (yes or no)

> Therefore the process doesn't terminate.

The problem is caused by the following change (introduced in May):

--- a/src/emacs.c
+++ b/src/emacs.c
@@ -385,6 +386,9 @@ fatal_error_signal (sig)
     {
       fatal_error_in_progress = 1;
 
+      if (sig == SIGTERM || sig == SIGHUP)
+        Fkill_emacs (make_number (sig));
+
       shut_down_emacs (sig, 0, Qnil);
     }
 

Re: Emacs daemon doesn't terminate on SIGTERM any more

[Jan Djärv]

Ulrich Mueller skrev 2010-10-23 00.51:
>>>>>> On Fri, 22 Oct 2010, Ulrich Mueller wrote:
>
>> Then I try to kill it from another terminal window:
>
>>     $ kill<pid of emacs process>
>
>> With Emacs 23.2 the process would properly terminate now. The BZR
>> version asks for interactive input in the open terminal frame instead:
>
>>      The current server still has clients; delete them? (yes or no)
>
>> Therefore the process doesn't terminate.
>
> The problem is caused by the following change (introduced in May):
>
> --- a/src/emacs.c
> +++ b/src/emacs.c
> @@ -385,6 +386,9 @@ fatal_error_signal (sig)
>       {
>         fatal_error_in_progress = 1;
>
> +      if (sig == SIGTERM || sig == SIGHUP)
> +        Fkill_emacs (make_number (sig));
> +
>         shut_down_emacs (sig, 0, Qnil);
>       }
>

This was done so that kill-emacs-hook is run.  For example, desktop.el will 
then save its desktop when logging out from Gnome/KDE.

What Emacs needs is some sort of notion of when user interaction is not OK. 
For SIGTERM it is almost never ok, as it might be the window 
manager/system/whatever that is shutting down.  Even desktop.el will try to 
ask questions sometimes (i.e. desktop-save is set but there is no desktop file 
yet), which is annoying.  Maybe introduce some interaction-ok-p predicate?

	Jan D.

Re: Emacs daemon doesn't terminate on SIGTERM any more

[Glenn Morris]

Jan Djärv wrote:

> This was done so that kill-emacs-hook is run.  For example, desktop.el
> will then save its desktop when logging out from Gnome/KDE.
>
> What Emacs needs is some sort of notion of when user interaction is
> not OK. For SIGTERM it is almost never ok, as it might be the window
> manager/system/whatever that is shutting down.  Even desktop.el will
> try to ask questions sometimes (i.e. desktop-save is set but there is
> no desktop file yet), which is annoying.  Maybe introduce some
> interaction-ok-p predicate?

Things that might ask questions should be on kill-emacs-query-functions,
not kill-emacs-hook, shouldn't they?

Re: Emacs daemon doesn't terminate on SIGTERM any more

[Jan Djärv]

Glenn Morris skrev 2010-10-23 20.40:
> Jan Djärv wrote:
>
>> This was done so that kill-emacs-hook is run.  For example, desktop.el
>> will then save its desktop when logging out from Gnome/KDE.
>>
>> What Emacs needs is some sort of notion of when user interaction is
>> not OK. For SIGTERM it is almost never ok, as it might be the window
>> manager/system/whatever that is shutting down.  Even desktop.el will
>> try to ask questions sometimes (i.e. desktop-save is set but there is
>> no desktop file yet), which is annoying.  Maybe introduce some
>> interaction-ok-p predicate?
>
> Things that might ask questions should be on kill-emacs-query-functions,
> not kill-emacs-hook, shouldn't they?

Not sure.  Things like desktop.el doesn't want to cancel the shutdown, just 
ask some questions.  The message that started this thread "The current server 
still has clients; delete them? (yes or no)" is from a function run from 
delete-frame-functions, which is run by delete-frame.

	Jan D.

Re: Emacs daemon doesn't terminate on SIGTERM any more

[Ulrich Mueller]

>>>>> On Sun, 24 Oct 2010, Jan Djärv wrote:

>>> What Emacs needs is some sort of notion of when user interaction
>>> is not OK. For SIGTERM it is almost never ok, as it might be the
>>> window manager/system/whatever that is shutting down. Even
>>> desktop.el will try to ask questions sometimes (i.e. desktop-save
>>> is set but there is no desktop file yet), which is annoying. Maybe
>>> introduce some interaction-ok-p predicate?
>> 
>> Things that might ask questions should be on
>> kill-emacs-query-functions, not kill-emacs-hook, shouldn't they?

> Not sure. Things like desktop.el doesn't want to cancel the
> shutdown, just ask some questions. The message that started this
> thread "The current server still has clients; delete them? (yes or
> no)" is from a function run from delete-frame-functions, which is
> run by delete-frame.

As far as I can see, the question is asked in function server-start.
Which in turn is called from kill-emacs-hook via function server-mode.

Also it doesn't make any difference if one answers the question with
yes or with no. The shutdown will proceed in either case.

Ulrich

Re: Emacs daemon doesn't terminate on SIGTERM any more

[Jan Djärv]

Ulrich Mueller skrev 2010-10-24 11.17:
>>>>>> On Sun, 24 Oct 2010, Jan Djärv wrote:
>
>>>> What Emacs needs is some sort of notion of when user interaction
>>>> is not OK. For SIGTERM it is almost never ok, as it might be the
>>>> window manager/system/whatever that is shutting down. Even
>>>> desktop.el will try to ask questions sometimes (i.e. desktop-save
>>>> is set but there is no desktop file yet), which is annoying. Maybe
>>>> introduce some interaction-ok-p predicate?
>>>
>>> Things that might ask questions should be on
>>> kill-emacs-query-functions, not kill-emacs-hook, shouldn't they?
>
>> Not sure. Things like desktop.el doesn't want to cancel the
>> shutdown, just ask some questions. The message that started this
>> thread "The current server still has clients; delete them? (yes or
>> no)" is from a function run from delete-frame-functions, which is
>> run by delete-frame.
>
> As far as I can see, the question is asked in function server-start.
> Which in turn is called from kill-emacs-hook via function server-mode.
>

Correct, I must have looked at the wrong function.

	Jan D.

> Also it doesn't make any difference if one answers the question with
> yes or with no. The shutdown will proceed in either case.
>
> Ulrich

Re: Emacs daemon doesn't terminate on SIGTERM any more

[Chong Yidong]

Jan Djärv <jan.h.d@swipnet.se> writes:

> This was done so that kill-emacs-hook is run.  For example, desktop.el
> will then save its desktop when logging out from Gnome/KDE.
>
> What Emacs needs is some sort of notion of when user interaction is
> not OK. For SIGTERM it is almost never ok, as it might be the window
> manager/system/whatever that is shutting down.  Even desktop.el will
> try to ask questions sometimes (i.e. desktop-save is set but there is
> no desktop file yet), which is annoying.  Maybe introduce some
> interaction-ok-p predicate?

How about it, in response to a kill signal, we do

  noninteractive = 1;

before calling Fkill_emacs?

Re: Emacs daemon doesn't terminate on SIGTERM any more

[Glenn Morris]

Jan Djärv wrote:

> Not sure.  Things like desktop.el doesn't want to cancel the shutdown,
> just ask some questions. 

I haven't looked at it properly, but I think maybe
kill-emacs-query-functions does the right thing, but has a bad
name/doc. Maybe it should be described as "List of functions to run
when killing Emacs that need to interact with the user. Additionally,
if any of them returns nil, killing Emacs is cancelled."

Re: Emacs daemon doesn't terminate on SIGTERM any more

[Jan Djärv]

Chong Yidong skrev 2010-10-24 18.06:
> Jan Djärv<jan.h.d@swipnet.se>  writes:
>
>> This was done so that kill-emacs-hook is run.  For example, desktop.el
>> will then save its desktop when logging out from Gnome/KDE.
>>
>> What Emacs needs is some sort of notion of when user interaction is
>> not OK. For SIGTERM it is almost never ok, as it might be the window
>> manager/system/whatever that is shutting down.  Even desktop.el will
>> try to ask questions sometimes (i.e. desktop-save is set but there is
>> no desktop file yet), which is annoying.  Maybe introduce some
>> interaction-ok-p predicate?
>
> How about it, in response to a kill signal, we do
>
>    noninteractive = 1;
>
> before calling Fkill_emacs?

That would be fine as long as code has something to look for.  We would have 
to go through those that out themselves into kill-emacs-hook and modify them, 
perhaps starting with server.el and desktop.el first.

	Jan D.