bug#12447: 24.1.50; Stuck in garbage collection on OS X

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Harald Hanche-Olsen]

This concerns emacs built with --with-ns from trunk on OS X.

To start with the symptom: Emacs seems to freeze, and spends a lot of
CPU time. Taking a sample with Activity monitor appears to indicate a
very deeply nested mark_object calls; from this I conclude that the
problem appears to happen in GC.

How to trigger the symptom: I always got it by trying to send an email
message with a large attachment. My mail client of choice is Mew, but
I suspect that any activity triggering garbage collection will also
trigger the bug.

After bisecting, I come to the conclusion that this revision is to
blame:

revno: 109470
committer: Jan D. <jan.h.d@swipnet.se>
branch nick: trunk
timestamp: Mon 2012-08-06 18:09:02 +0200
message:
  * keyboard.c (timer_check_2): Add break so timer_check returns next timeout.

That revision only adds a single line of code. I also find that the
bug disappears if I patch the current tip of trunk (revision 110013)
as follows, which undoes revision 109470:

=== modified file 'src/keyboard.c'
--- src/keyboard.c	2012-09-13 02:21:28 +0000
+++ src/keyboard.c	2012-09-14 19:12:15 +0000
@@ -4484,7 +4484,6 @@
 	    }
 
 	  nexttime = make_emacs_time (0, 0);
-          break;
 	}
       else
 	/* When we encounter a timer that is still waiting,


I suppose the break was added for a good reason, so this is probably
not the right thing. But it does cure the present problem.

(This bug is also discussed on the emacs-devel list, in the thread
titled "Emacs seems awfully unstable on OS X lately".)

There seems to be a second bug, introduced later, that causes crashes
(as opposed to hangs). I mention this because the second bug could
interfere with attempts to work on this one. I will endeavour to track
down the other bug as well.

- Harald

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Jan Djärv]

Hello.

I think this is probably the same bug as in #12326.  I.e. a timer gets added over and over again.
Please consider how to fix the original problem as described in http://debbugs.gnu.org/cgi/bugreport.cgi?bug=12326#11 before removing the break statement.

	Jan D.

14 sep 2012 kl. 23:08 skrev Harald Hanche-Olsen <hanche@math.ntnu.no>:

> This concerns emacs built with --with-ns from trunk on OS X.
> 
> To start with the symptom: Emacs seems to freeze, and spends a lot of
> CPU time. Taking a sample with Activity monitor appears to indicate a
> very deeply nested mark_object calls; from this I conclude that the
> problem appears to happen in GC.
> 
> How to trigger the symptom: I always got it by trying to send an email
> message with a large attachment. My mail client of choice is Mew, but
> I suspect that any activity triggering garbage collection will also
> trigger the bug.
> 
> After bisecting, I come to the conclusion that this revision is to
> blame:
> 
> revno: 109470
> committer: Jan D. <jan.h.d@swipnet.se>
> branch nick: trunk
> timestamp: Mon 2012-08-06 18:09:02 +0200
> message:
>  * keyboard.c (timer_check_2): Add break so timer_check returns next timeout.
> 
> That revision only adds a single line of code. I also find that the
> bug disappears if I patch the current tip of trunk (revision 110013)
> as follows, which undoes revision 109470:
> 
> === modified file 'src/keyboard.c'
> --- src/keyboard.c	2012-09-13 02:21:28 +0000
> +++ src/keyboard.c	2012-09-14 19:12:15 +0000
> @@ -4484,7 +4484,6 @@
> 	    }
> 
> 	  nexttime = make_emacs_time (0, 0);
> -          break;
> 	}
>       else
> 	/* When we encounter a timer that is still waiting,
> 
> 
> I suppose the break was added for a good reason, so this is probably
> not the right thing. But it does cure the present problem.
> 
> (This bug is also discussed on the emacs-devel list, in the thread
> titled "Emacs seems awfully unstable on OS X lately".)
> 
> There seems to be a second bug, introduced later, that causes crashes
> (as opposed to hangs). I mention this because the second bug could
> interfere with attempts to work on this one. I will endeavour to track
> down the other bug as well.
> 
> - Harald

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Harald Hanche-Olsen]

[Jan Djärv <jan.h.d@swipnet.se> (2012-09-15 09:55:55 UTC)]

> I think this is probably the same bug as in #12326.  I.e. a timer gets added over and over again.
> Please consider how to fix the original problem as described in http://debbugs.gnu.org/cgi/bugreport.cgi?bug=12326#11 before removing the break statement.

Thanks for the pointer. There are certainly timers involved in the code that runs when I get the hangs; I'll review it and see if that is where the problem lies.

However, I am not sure exactly what to look for. Or, more precisely, what is the undefined behaviour in js2-mode that you complain about it bug 12326?

On a side note, I see that only a few timer functions are described in the elisp manual. It appears one has to read timer.el to find out how timers work at a lower level. Or did I miss something?

- Harald

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Sat, 15 Sep 2012 13:07:10 +0200 (CEST)
> From: Harald Hanche-Olsen <hanche@math.ntnu.no>
> Cc: 12447@debbugs.gnu.org
> 
> It appears one has to read timer.el to find out how timers work at a lower level. Or did I miss something?

No, the real story of how timers work is in process.c.  See the
function wait_reading_process_output there, which runs the timers via
the call to timer_check around line 4444; time_check and its
subroutines are in keyboard.c.

timer.el is just the Lisp interface to that code, used to create,
delete, and manage timer objects.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Harald Hanche-Olsen]

[Eli Zaretskii <eliz@gnu.org> (2012-09-15 12:31:50 UTC)]

> No, the real story of how timers work is in process.c.  See the
> function wait_reading_process_output there, which runs the timers via
> the call to timer_check around line 4444; time_check and its
> subroutines are in keyboard.c.
> 
> timer.el is just the Lisp interface to that code, used to create,
> delete, and manage timer objects.

Okay, I'll study it; but it seems ludicrous that you need to read C
code in order to use timers from elisp without causing emacs to hang.
I almost feel like filing a documentation bug report. But I'll try to
gain a better understanding of the issue before considering that.

- Harald

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Sat, 15 Sep 2012 15:19:09 +0200 (CEST)
> Cc: jan.h.d@swipnet.se, 12447@debbugs.gnu.org
> From: Harald Hanche-Olsen <hanche@math.ntnu.no>
> 
> [Eli Zaretskii <eliz@gnu.org> (2012-09-15 12:31:50 UTC)]
> 
> > No, the real story of how timers work is in process.c.  See the
> > function wait_reading_process_output there, which runs the timers via
> > the call to timer_check around line 4444; time_check and its
> > subroutines are in keyboard.c.
> > 
> > timer.el is just the Lisp interface to that code, used to create,
> > delete, and manage timer objects.
> 
> Okay, I'll study it; but it seems ludicrous that you need to read C
> code in order to use timers from elisp without causing emacs to hang.

??? AFAIU, you are trying to track a bug, not use the timers.  You
specifically asked about "how the timers work at a lower level".

Using timers does not involve any reading of C code; just read the
"Timers" and "Idle timers" nodes of the ELisp manual, and Bob's your
uncle.  No low-level knowledge is needed just for using timers.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Harald Hanche-Olsen]

[Eli Zaretskii <eliz@gnu.org> (2012-09-15 13:56:14 UTC)]

> ??? AFAIU, you are trying to track a bug, not use the timers.  You
> specifically asked about "how the timers work at a lower level".
> 
> Using timers does not involve any reading of C code; just read the
> "Timers" and "Idle timers" nodes of the ELisp manual, and Bob's your
> uncle.  No low-level knowledge is needed just for using timers.

Ah. I was forming my impression by looking at the code quoted in the
discussion of bug#12326, which seems to be using lower level stuff. So
I thought I had to understand that. But the timer code in mew seems to
only use the interface described in the elisp manual.

But yes, I am chasing a bug, not trying to use timers. Which seems to
require understanding the underlying mechanism, as otherwise I have no
idea what I need to be looking for. After all, the code works fine
most of the time, but revision 109470 broke it for some strange
reason. And this seems to imply that you do need to understand the
lower levels in order to use timers safely. Which is bad, if true.

For example: Is it okay to set a timer in a timer callback? Example
code in the "Idle timers" node in the elisp manual seems to say so.

Thanks for your patience with my many misunderstandings. I want to
track this problem down, but I am still fumbling in the dark.

- Harald

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Sat, 15 Sep 2012 16:23:53 +0200 (CEST)
> Cc: jan.h.d@swipnet.se, 12447@debbugs.gnu.org
> From: Harald Hanche-Olsen <hanche@math.ntnu.no>
> 
> For example: Is it okay to set a timer in a timer callback?

I don't see why not.  Setting up a timer just creates a Lisp object
and adds it to the list of timers.  The timer will be run the next
time the low-level mechanism, which is part of the Emacs exec loop,
determines that it's ripe.

> Thanks for your patience with my many misunderstandings. I want to
> track this problem down, but I am still fumbling in the dark.

You are welcome.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Jan Djärv]

15 sep 2012 kl. 13:07 skrev Harald Hanche-Olsen <hanche@math.ntnu.no>:

> [Jan Djärv <jan.h.d@swipnet.se> (2012-09-15 09:55:55 UTC)]
> 
>> I think this is probably the same bug as in #12326.  I.e. a timer gets added over and over again.
>> Please consider how to fix the original problem as described in http://debbugs.gnu.org/cgi/bugreport.cgi?bug=12326#11 before removing the break statement.
> 
> Thanks for the pointer. There are certainly timers involved in the code that runs when I get the hangs; I'll review it and see if that is where the problem lies.
> 
> However, I am not sure exactly what to look for. Or, more precisely, what is the undefined behaviour in js2-mode that you complain about it bug 12326?
> 

The undefined behaviour was before adding the break, timers could run at a time in the future that depended on other timers.  The defined behaviour is with the break, timers run when they should.

I suspect there is some bug in the C code when adding a timer from within a timer callback.

> On a side note, I see that only a few timer functions are described in the elisp manual. It appears one has to read timer.el to find out how timers work at a lower level. Or did I miss something?

As Eli has answered, the timers are handeled in C as a list sorted by time when the timers are to be run.  I.e. the next timer to run is at the head of the list.

	Jan D.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

Eli Zaretskii <eliz@gnu.org> writes:

 >> Date: Sat, 15 Sep 2012 16:23:53 +0200 (CEST)
 >> Cc: jan.h.d@swipnet.se, 12447@debbugs.gnu.org
 >> From: Harald Hanche-Olsen <hanche@math.ntnu.no>
 >>
 >> For example: Is it okay to set a timer in a timer callback?
 >
 > I don't see why not.  Setting up a timer just creates a Lisp object
 > and adds it to the list of timers.  The timer will be run the next
 > time the low-level mechanism, which is part of the Emacs exec loop,
 > determines that it's ripe.

In js2-mode's case, the problem is that 'run-with-idle-timer' makes the
created timer run now, not the "next time".  Here's an example:

   (defvar counter 0)

   (defun foo ()
     (message (format  "foo %s" counter))
     (incf counter)
     (run-with-idle-timer 1 nil #'foo))

   (foo)

I'd expect that either timer would fire once every second (as long as
I'm not touching my keyboard), or at least stop firing when I do touch
my keyboard (seeing as otherwise Emacs is idle), but instead I just see
the timer firing many times a second, the counter runs in the message
area, and Emacs doesn't respond to any commands.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Sun, 16 Sep 2012 13:15:52 +0400
> From: Dmitry Gutov <dgutov@yandex.ru>
> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
> 
> Eli Zaretskii <eliz@gnu.org> writes:
> 
>  >> Date: Sat, 15 Sep 2012 16:23:53 +0200 (CEST)
>  >> Cc: jan.h.d@swipnet.se, 12447@debbugs.gnu.org
>  >> From: Harald Hanche-Olsen <hanche@math.ntnu.no>
>  >>
>  >> For example: Is it okay to set a timer in a timer callback?
>  >
>  > I don't see why not.  Setting up a timer just creates a Lisp object
>  > and adds it to the list of timers.  The timer will be run the next
>  > time the low-level mechanism, which is part of the Emacs exec loop,
>  > determines that it's ripe.
> 
> In js2-mode's case, the problem is that 'run-with-idle-timer' makes the
> created timer run now, not the "next time".  Here's an example:
> 
>    (defvar counter 0)
> 
>    (defun foo ()
>      (message (format  "foo %s" counter))
>      (incf counter)
>      (run-with-idle-timer 1 nil #'foo))
> 
>    (foo)

The code above does not run the timer, it just schedules it to run
after at least 1 sec of idleness time.

> I'd expect that either timer would fire once every second (as long as
> I'm not touching my keyboard), or at least stop firing when I do touch
> my keyboard (seeing as otherwise Emacs is idle), but instead I just see
> the timer firing many times a second, the counter runs in the message
> area, and Emacs doesn't respond to any commands.

Then there's a bug, because an idle timer should only fire when
there's no other input.  If there's keyboard input, Emacs should
process it first.

I don't think the bug is related to the fact that the timer handler
re-schedules itself.  That is something many timers do.  There's
something else at work here, and that something is most probably on
the C level.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 16.09.2012 14:31, Eli Zaretskii wrote:
>> Date: Sun, 16 Sep 2012 13:15:52 +0400
>> From: Dmitry Gutov <dgutov@yandex.ru>
>> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
>>
>> Eli Zaretskii <eliz@gnu.org> writes:
>>
>>   >> Date: Sat, 15 Sep 2012 16:23:53 +0200 (CEST)
>>   >> Cc: jan.h.d@swipnet.se, 12447@debbugs.gnu.org
>>   >> From: Harald Hanche-Olsen <hanche@math.ntnu.no>
>>   >>
>>   >> For example: Is it okay to set a timer in a timer callback?
>>   >
>>   > I don't see why not.  Setting up a timer just creates a Lisp object
>>   > and adds it to the list of timers.  The timer will be run the next
>>   > time the low-level mechanism, which is part of the Emacs exec loop,
>>   > determines that it's ripe.
>>
>> In js2-mode's case, the problem is that 'run-with-idle-timer' makes the
>> created timer run now, not the "next time".  Here's an example:
>>
>>     (defvar counter 0)
>>
>>     (defun foo ()
>>       (message (format  "foo %s" counter))
>>       (incf counter)
>>       (run-with-idle-timer 1 nil #'foo))
>>
>>     (foo)
>
> The code above does not run the timer, it just schedules it to run
> after at least 1 sec of idleness time.

Yes. And 1 second later, Emacs effectively freezes (while still 
continuing to show the increasing counter in the message area).

>> I'd expect that either timer would fire once every second (as long as
>> I'm not touching my keyboard), or at least stop firing when I do touch
>> my keyboard (seeing as otherwise Emacs is idle), but instead I just see
>> the timer firing many times a second, the counter runs in the message
>> area, and Emacs doesn't respond to any commands.
>
> Then there's a bug, because an idle timer should only fire when
> there's no other input.  If there's keyboard input, Emacs should
> process it first.
>
> I don't think the bug is related to the fact that the timer handler
> re-schedules itself.  That is something many timers do.  There's
> something else at work here, and that something is most probably on
> the C level.

Like I wrote in 12326, AFAICT, the problem is that timer_check_2 doesn't 
at any point check that Emacs is still idle. When run-with-idle-timer 
calls (timer-activate-when-idle timer t), the new timer is added to the 
list, timer_check_2 reaches is and runs it immediately because 
'timer_idleness_start_time' still has the same value.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Sun, 16 Sep 2012 14:44:30 +0400
> From: Dmitry Gutov <dgutov@yandex.ru>
> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
> 
> Like I wrote in 12326, AFAICT, the problem is that timer_check_2 doesn't 
> at any point check that Emacs is still idle. When run-with-idle-timer 
> calls (timer-activate-when-idle timer t), the new timer is added to the 
> list, timer_check_2 reaches is and runs it immediately because 
> 'timer_idleness_start_time' still has the same value.

If that is the problem, then perhaps having timer_check_2 work on a
copy of the list would solve the problem.  Did you try that?

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 16.09.2012 15:53, Eli Zaretskii wrote:
>> Date: Sun, 16 Sep 2012 14:44:30 +0400
>> From: Dmitry Gutov <dgutov@yandex.ru>
>> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
>>
>> Like I wrote in 12326, AFAICT, the problem is that timer_check_2 doesn't
>> at any point check that Emacs is still idle. When run-with-idle-timer
>> calls (timer-activate-when-idle timer t), the new timer is added to the
>> list, timer_check_2 reaches is and runs it immediately because
>> 'timer_idleness_start_time' still has the same value.
>
> If that is the problem, then perhaps having timer_check_2 work on a
> copy of the list would solve the problem.  Did you try that?

I'm no C programmer, so I didn't try to fix it in C code. How would I 
make a copy of a list there?

But no, it probably won't: the "guilty" commit made timer_check_2 
actually return 0 after a timer fires (keeping true to the comment above 
it), so a local copy would serve no purpose.

Assuming it would've worked, though, wouldn't it make the second 
argument to 'timer-activate-when-idle' useless?

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Sun, 16 Sep 2012 16:07:07 +0400
> From: Dmitry Gutov <dgutov@yandex.ru>
> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
> 
> >> Like I wrote in 12326, AFAICT, the problem is that timer_check_2 doesn't
> >> at any point check that Emacs is still idle. When run-with-idle-timer
> >> calls (timer-activate-when-idle timer t), the new timer is added to the
> >> list, timer_check_2 reaches is and runs it immediately because
> >> 'timer_idleness_start_time' still has the same value.
> >
> > If that is the problem, then perhaps having timer_check_2 work on a
> > copy of the list would solve the problem.  Did you try that?
> 
> I'm no C programmer, so I didn't try to fix it in C code. How would I 
> make a copy of a list there?

Using the Fcopy_sequence function, I'd think.  Use it at the beginning
of the function to set the value of 'idle_timers', instead of this
line:

    idle_timers = Vtimer_idle_list;

> But no, it probably won't: the "guilty" commit made timer_check_2 
> actually return 0 after a timer fires (keeping true to the comment above 
> it), so a local copy would serve no purpose.

timer_check_2 indeed returns, but then timer_check will call it again,
because it continues calling timer_check_2 in a loop, until there's no
ripe timer.

> Assuming it would've worked, though, wouldn't it make the second 
> argument to 'timer-activate-when-idle' useless?

It isn't useless, it just means a slight delay in the "now" part of
its doc string.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 16.09.2012 16:39, Eli Zaretskii wrote:
>> Date: Sun, 16 Sep 2012 16:07:07 +0400
>> From: Dmitry Gutov <dgutov@yandex.ru>
>> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
>>
>>>> Like I wrote in 12326, AFAICT, the problem is that timer_check_2 doesn't
>>>> at any point check that Emacs is still idle. When run-with-idle-timer
>>>> calls (timer-activate-when-idle timer t), the new timer is added to the
>>>> list, timer_check_2 reaches is and runs it immediately because
>>>> 'timer_idleness_start_time' still has the same value.
>>>
>>> If that is the problem, then perhaps having timer_check_2 work on a
>>> copy of the list would solve the problem.  Did you try that?
>>
>> I'm no C programmer, so I didn't try to fix it in C code. How would I
>> make a copy of a list there?
>
> Using the Fcopy_sequence function, I'd think.  Use it at the beginning
> of the function to set the value of 'idle_timers', instead of this
> line:
>
>      idle_timers = Vtimer_idle_list;

Done that, recompiled, no difference in the example a sent previously. 
Exactly because, I think, of the control flow you describe below:

>> But no, it probably won't: the "guilty" commit made timer_check_2
>> actually return 0 after a timer fires (keeping true to the comment above
>> it), so a local copy would serve no purpose.
>
> timer_check_2 indeed returns, but then timer_check will call it again,
> because it continues calling timer_check_2 in a loop, until there's no
> ripe timer.

Each time timer_check_2 is called, a new copy would be made from the 
idle timers list, so the newly created timer would be reached during the 
same call to 'timer_check'.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Sun, 16 Sep 2012 17:25:35 +0400
> From: Dmitry Gutov <dgutov@yandex.ru>
> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
> 
> >      idle_timers = Vtimer_idle_list;
> 
> Done that, recompiled, no difference in the example a sent previously. 
> Exactly because, I think, of the control flow you describe below:
> 
> >> But no, it probably won't: the "guilty" commit made timer_check_2
> >> actually return 0 after a timer fires (keeping true to the comment above
> >> it), so a local copy would serve no purpose.
> >
> > timer_check_2 indeed returns, but then timer_check will call it again,
> > because it continues calling timer_check_2 in a loop, until there's no
> > ripe timer.
> 
> Each time timer_check_2 is called, a new copy would be made from the 
> idle timers list, so the newly created timer would be reached during the 
> same call to 'timer_check'.

But the first time timer_check_2 encounters an un-ripe timer, or gets
to the end of the list of timers, it returns a non-zero 'nexttime'
value, which should break the loop in timer_check.  Why isn't this
happening?

Perhaps the problem is that the value of 'difference' is not
initialized:

  while (CONSP (timers) || CONSP (idle_timers))
    {
      Lisp_Object timer = Qnil, idle_timer = Qnil;
      EMACS_TIME timer_time, idle_timer_time;
      EMACS_TIME difference;   <<<<<<<<<<<<<<<<<<<<<<<<<<<<

and then never set to any specific value, until here:

      else
	/* When we encounter a timer that is still waiting,
	   return the amount of time to wait before it is ripe.  */
	{
	  UNGCPRO;
	  return difference;
	}

which causes us return garbage, potentially zero, to timer_check.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 16.09.2012 17:47, Eli Zaretskii wrote:
>> Date: Sun, 16 Sep 2012 17:25:35 +0400
>> From: Dmitry Gutov <dgutov@yandex.ru>
>> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
>>
>>>       idle_timers = Vtimer_idle_list;
>>
>> Done that, recompiled, no difference in the example a sent previously.
>> Exactly because, I think, of the control flow you describe below:
>>
>>>> But no, it probably won't: the "guilty" commit made timer_check_2
>>>> actually return 0 after a timer fires (keeping true to the comment above
>>>> it), so a local copy would serve no purpose.
>>>
>>> timer_check_2 indeed returns, but then timer_check will call it again,
>>> because it continues calling timer_check_2 in a loop, until there's no
>>> ripe timer.
>>
>> Each time timer_check_2 is called, a new copy would be made from the
>> idle timers list, so the newly created timer would be reached during the
>> same call to 'timer_check'.
>
> But the first time timer_check_2 encounters an un-ripe timer, or gets
> to the end of the list of timers, it returns a non-zero 'nexttime'
> value, which should break the loop in timer_check.  Why isn't this
> happening?

Maybe because the new timer is at the top of the list, and it's 
considered ripe?

> Perhaps the problem is that the value of 'difference' is not
> initialized:
>
>    while (CONSP (timers) || CONSP (idle_timers))
>      {
>        Lisp_Object timer = Qnil, idle_timer = Qnil;
>        EMACS_TIME timer_time, idle_timer_time;
>        EMACS_TIME difference;   <<<<<<<<<<<<<<<<<<<<<<<<<<<<
>
> and then never set to any specific value, until here:
>
>        else
> 	/* When we encounter a timer that is still waiting,
> 	   return the amount of time to wait before it is ripe.  */
> 	{
> 	  UNGCPRO;
> 	  return difference;
> 	}
>
> which causes us return garbage, potentially zero, to timer_check.

It's assigned to, though. When we encounter a timer that's not yet ripe.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Sun, 16 Sep 2012 18:25:50 +0400
> From: Dmitry Gutov <dgutov@yandex.ru>
> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
> 
> > Perhaps the problem is that the value of 'difference' is not
> > initialized:
> >
> >    while (CONSP (timers) || CONSP (idle_timers))
> >      {
> >        Lisp_Object timer = Qnil, idle_timer = Qnil;
> >        EMACS_TIME timer_time, idle_timer_time;
> >        EMACS_TIME difference;   <<<<<<<<<<<<<<<<<<<<<<<<<<<<
> >
> > and then never set to any specific value, until here:
> >
> >        else
> > 	/* When we encounter a timer that is still waiting,
> > 	   return the amount of time to wait before it is ripe.  */
> > 	{
> > 	  UNGCPRO;
> > 	  return difference;
> > 	}
> >
> > which causes us return garbage, potentially zero, to timer_check.
> 
> It's assigned to, though. When we encounter a timer that's not yet ripe.

What if all of them are ripe?

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 16.09.2012 18:54, Eli Zaretskii wrote:
>> Date: Sun, 16 Sep 2012 18:25:50 +0400
>> From: Dmitry Gutov <dgutov@yandex.ru>
>> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
>>
>>> Perhaps the problem is that the value of 'difference' is not
>>> initialized:
>>>
>>>     while (CONSP (timers) || CONSP (idle_timers))
>>>       {
>>>         Lisp_Object timer = Qnil, idle_timer = Qnil;
>>>         EMACS_TIME timer_time, idle_timer_time;
>>>         EMACS_TIME difference;   <<<<<<<<<<<<<<<<<<<<<<<<<<<<
>>>
>>> and then never set to any specific value, until here:
>>>
>>>         else
>>> 	/* When we encounter a timer that is still waiting,
>>> 	   return the amount of time to wait before it is ripe.  */
>>> 	{
>>> 	  UNGCPRO;
>>> 	  return difference;
>>> 	}
>>>
>>> which causes us return garbage, potentially zero, to timer_check.
>>
>> It's assigned to, though. When we encounter a timer that's not yet ripe.
>
> What if all of them are ripe?

I don't see the problem. The first timer is ripe? Fire it and return 
'nexttime'. Otherwise, return 'difference', which now has been assigned 
a value.
If we've reached the end of the list, again return 'nexttime', which is 
initialized with invalid_emacs_time () at the beginning of timer_check_2.

Anyway, I think the immediate problem is that the newly created timer 
can be considered ripe.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Sun, 16 Sep 2012 19:56:11 +0400
> From: Dmitry Gutov <dgutov@yandex.ru>
> Cc: 12447@debbugs.gnu.org, hanche@math.ntnu.no
> 
> On 16.09.2012 18:54, Eli Zaretskii wrote:
> >> Date: Sun, 16 Sep 2012 18:25:50 +0400
> >> From: Dmitry Gutov <dgutov@yandex.ru>
> >> CC: hanche@math.ntnu.no, 12447@debbugs.gnu.org
> >>
> >>> Perhaps the problem is that the value of 'difference' is not
> >>> initialized:
> >>>
> >>>     while (CONSP (timers) || CONSP (idle_timers))
> >>>       {
> >>>         Lisp_Object timer = Qnil, idle_timer = Qnil;
> >>>         EMACS_TIME timer_time, idle_timer_time;
> >>>         EMACS_TIME difference;   <<<<<<<<<<<<<<<<<<<<<<<<<<<<
> >>>
> >>> and then never set to any specific value, until here:
> >>>
> >>>         else
> >>> 	/* When we encounter a timer that is still waiting,
> >>> 	   return the amount of time to wait before it is ripe.  */
> >>> 	{
> >>> 	  UNGCPRO;
> >>> 	  return difference;
> >>> 	}
> >>>
> >>> which causes us return garbage, potentially zero, to timer_check.
> >>
> >> It's assigned to, though. When we encounter a timer that's not yet ripe.
> >
> > What if all of them are ripe?
> 
> I don't see the problem. The first timer is ripe? Fire it and return 
> 'nexttime'. Otherwise, return 'difference', which now has been assigned 
> a value.
> If we've reached the end of the list, again return 'nexttime', which is 
> initialized with invalid_emacs_time () at the beginning of timer_check_2.
> 
> Anyway, I think the immediate problem is that the newly created timer 
> can be considered ripe.

The patch below makes your simplified recipe, viz.:

  (defvar counter 0)

  (defun foo ()
    (message (format  "foo %s" counter))
    (setq counter (1+ counter))
    (run-with-idle-timer 1 nil #'foo))
  (foo)

"work" without locking up Emacs.  "Work" in the sense that the timer
is run and increments the counter, but keyboard input is still
accepted, and causes 1-sec break in the idle timer invocation.  What
does NOT happen is the once-per-second invocation of the idle timer:
as long as there's no other input, the idle timer runs much more
frequently.  But I think this is expected, since the call to
run-with-idle-timer above explicitly asks to be run immediately.

Can you see if these changes also make js2-mode work as expected?

Jan, can you test whether this patch still keeps your two-timers
recipe working?  If it does, I think I should commit the changes
below, because they avoid locking up Emacs by a timer that repeatedly
reinvokes itself.


=== modified file 'src/keyboard.c'
--- src/keyboard.c	2012-09-16 21:43:55 +0000
+++ src/keyboard.c	2012-09-18 14:58:24 +0000
@@ -4334,25 +4334,18 @@ decode_timer (Lisp_Object timer, EMACS_T
    should be done.  */
 
 static EMACS_TIME
-timer_check_2 (void)
+timer_check_2 (Lisp_Object timers, Lisp_Object idle_timers)
 {
   EMACS_TIME nexttime;
   EMACS_TIME now;
   EMACS_TIME idleness_now;
-  Lisp_Object timers, idle_timers, chosen_timer;
-  struct gcpro gcpro1, gcpro2, gcpro3;
+  Lisp_Object chosen_timer;
+  struct gcpro gcpro1;
 
   nexttime = invalid_emacs_time ();
 
-  /* Always consider the ordinary timers.  */
-  timers = Vtimer_list;
-  /* Consider the idle timers only if Emacs is idle.  */
-  if (EMACS_TIME_VALID_P (timer_idleness_start_time))
-    idle_timers = Vtimer_idle_list;
-  else
-    idle_timers = Qnil;
   chosen_timer = Qnil;
-  GCPRO3 (timers, idle_timers, chosen_timer);
+  GCPRO1 (chosen_timer);
 
   /* First run the code that was delayed.  */
   while (CONSP (pending_funcalls))
@@ -4501,13 +4494,30 @@ EMACS_TIME
 timer_check (void)
 {
   EMACS_TIME nexttime;
+  Lisp_Object timers, idle_timers;
+  struct gcpro gcpro1, gcpro2;
+
+  /* We use copies of the timers' lists to allow a timer to add itself
+     again, without locking up Emacs if the newly added timer is
+     already ripe when added.  */
+
+  /* Always consider the ordinary timers.  */
+  timers = Fcopy_sequence (Vtimer_list);
+  /* Consider the idle timers only if Emacs is idle.  */
+  if (EMACS_TIME_VALID_P (timer_idleness_start_time))
+    idle_timers = Fcopy_sequence (Vtimer_idle_list);
+  else
+    idle_timers = Qnil;
+
+  GCPRO2 (timers, idle_timers);
 
   do
     {
-      nexttime = timer_check_2 ();
+      nexttime = timer_check_2 (timers, idle_timers);
     }
   while (EMACS_SECS (nexttime) == 0 && EMACS_NSECS (nexttime) == 0);
 
+  UNGCPRO;
   return nexttime;
 }
 

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Jan Djärv]

Hello.

18 sep 2012 kl. 17:05 skrev Eli Zaretskii <eliz@gnu.org>:

> The patch below makes your simplified recipe, viz.:
> 
>  (defvar counter 0)
> 
>  (defun foo ()
>    (message (format  "foo %s" counter))
>    (setq counter (1+ counter))
>    (run-with-idle-timer 1 nil #'foo))
>  (foo)
> 
> "work" without locking up Emacs.  "Work" in the sense that the timer
> is run and increments the counter, but keyboard input is still
> accepted, and causes 1-sec break in the idle timer invocation.  What
> does NOT happen is the once-per-second invocation of the idle timer:
> as long as there's no other input, the idle timer runs much more
> frequently.  But I think this is expected, since the call to
> run-with-idle-timer above explicitly asks to be run immediately.

> 
> Can you see if these changes also make js2-mode work as expected?
> 
> Jan, can you test whether this patch still keeps your two-timers
> recipe working?  If it does, I think I should commit the changes
> below, because they avoid locking up Emacs by a timer that repeatedly
> reinvokes itself.


It does.

	Jan D.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> From: Jan Djärv <jan.h.d@swipnet.se>
> Date: Tue, 18 Sep 2012 19:29:28 +0200
> Cc: Dmitry Gutov <dgutov@yandex.ru>,
>  12447@debbugs.gnu.org,
>  hanche@math.ntnu.no
> 
> > Jan, can you test whether this patch still keeps your two-timers
> > recipe working?  If it does, I think I should commit the changes
> > below, because they avoid locking up Emacs by a timer that repeatedly
> > reinvokes itself.
> 
> 
> It does.

Thanks.  Assuming js2-mode also works, I will wait for a few days to
hear possible objections and comments, and commit then.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 18.09.2012 19:05, Eli Zaretskii wrote:
> The patch below makes your simplified recipe, viz.:
>
>    (defvar counter 0)
>
>    (defun foo ()
>      (message (format  "foo %s" counter))
>      (setq counter (1+ counter))
>      (run-with-idle-timer 1 nil #'foo))
>    (foo)
>
> "work" without locking up Emacs.  "Work" in the sense that the timer
> is run and increments the counter, but keyboard input is still
> accepted, and causes 1-sec break in the idle timer invocation.  What
> does NOT happen is the once-per-second invocation of the idle timer:
> as long as there's no other input, the idle timer runs much more
> frequently.  But I think this is expected, since the call to
> run-with-idle-timer above explicitly asks to be run immediately.

I think this behavior makes sense, too.

> Can you see if these changes also make js2-mode work as expected?

They do, thank you. I'll keep the workaround, though, as it reportedly 
also fixes the long-standing OS X freeze bug.

--Dmitry

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Wed, 19 Sep 2012 04:27:07 +0400
> From: Dmitry Gutov <dgutov@yandex.ru>
> CC: Jan Djärv <jan.h.d@swipnet.se>, 
>  12447@debbugs.gnu.org, hanche@math.ntnu.no
> 
> > Can you see if these changes also make js2-mode work as expected?
> 
> They do, thank you.

Thanks for testing.

> I'll keep the workaround, though, as it reportedly also fixes the
> long-standing OS X freeze bug.

What workaround is that?

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Harald Hanche-Olsen]

[Eli Zaretskii <eliz@gnu.org> (2012-09-19 02:54:39 UTC)]

> > Date: Wed, 19 Sep 2012 04:27:07 +0400
> > From: Dmitry Gutov <dgutov@yandex.ru>
> > CC: Jan Djärv <jan.h.d@swipnet.se>, 
> >  12447@debbugs.gnu.org, hanche@math.ntnu.no
> > 
> > I'll keep the workaround, though, as it reportedly also fixes the
> > long-standing OS X freeze bug.
> 
> What workaround is that?

Perhaps he meant backing out of 109470.

FWIW, I am now running with your patch applied, and 109470 reinstated,
and sending messages with big attachments now works without a hitch.

- Harald

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 19.09.2012 6:54, Eli Zaretskii wrote:
>> I'll keep the workaround, though, as it reportedly also fixes the
>> long-standing OS X freeze bug.
>
> What workaround is that?

The one I posted in #12326: calling timer-activate-when-idle with nil 
DONT-WAIT argument.

https://github.com/mooz/js2-mode/commit/b02f4a0d72d0e2087038fe891e2580c4505415ef

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Wed, 19 Sep 2012 09:41:10 +0200 (CEST)
> Cc: dgutov@yandex.ru, jan.h.d@swipnet.se, 12447@debbugs.gnu.org
> From: Harald Hanche-Olsen <hanche@math.ntnu.no>
> 
> [Eli Zaretskii <eliz@gnu.org> (2012-09-19 02:54:39 UTC)]
> 
> > > Date: Wed, 19 Sep 2012 04:27:07 +0400
> > > From: Dmitry Gutov <dgutov@yandex.ru>
> > > CC: Jan Djärv <jan.h.d@swipnet.se>, 
> > >  12447@debbugs.gnu.org, hanche@math.ntnu.no
> > > 
> > > I'll keep the workaround, though, as it reportedly also fixes the
> > > long-standing OS X freeze bug.
> > 
> > What workaround is that?
> 
> Perhaps he meant backing out of 109470.

I don't think it should be backed out.

> FWIW, I am now running with your patch applied, and 109470 reinstated,
> and sending messages with big attachments now works without a hitch.

Thanks.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Wed, 19 Sep 2012 13:54:39 +0400
> From: Dmitry Gutov <dgutov@yandex.ru>
> CC: jan.h.d@swipnet.se, 12447@debbugs.gnu.org, hanche@math.ntnu.no
> 
> On 19.09.2012 6:54, Eli Zaretskii wrote:
> >> I'll keep the workaround, though, as it reportedly also fixes the
> >> long-standing OS X freeze bug.
> >
> > What workaround is that?
> 
> The one I posted in #12326: calling timer-activate-when-idle with nil 
> DONT-WAIT argument.

I don't think that's a workaround.  I think that's what js2 should do,
if (AFAIU) it wants the timer handler be invoked once, after Emacs has
been idle for more than 1 sec.

> https://github.com/mooz/js2-mode/commit/b02f4a0d72d0e2087038fe891e2580c4505415ef

Btw, in this line:

  (timer-activate-when-idle timer nil)))

I think you can lose the last argument, since it is optional anyway.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 19.09.2012 19:24, Eli Zaretskii wrote:
>> Date: Wed, 19 Sep 2012 13:54:39 +0400
>> From: Dmitry Gutov <dgutov@yandex.ru>
>> CC: jan.h.d@swipnet.se, 12447@debbugs.gnu.org, hanche@math.ntnu.no
>>
>> On 19.09.2012 6:54, Eli Zaretskii wrote:
>>>> I'll keep the workaround, though, as it reportedly also fixes the
>>>> long-standing OS X freeze bug.
>>>
>>> What workaround is that?
>>
>> The one I posted in #12326: calling timer-activate-when-idle with nil
>> DONT-WAIT argument.
>
> I don't think that's a workaround.  I think that's what js2 should do,
> if (AFAIU) it wants the timer handler be invoked once, after Emacs has
> been idle for more than 1 sec.

Maybe, but run-with-idle-timer calls it with DONT-WAIT t, and 
js2-mode-reset-timer is only supposed to be called in response to some 
user action.
So with your patch, it works effectively the same either way.

By the way, here's what run-with-idle-timer docstring says:
"Perform an action the next time Emacs is idle for SECS seconds."

Shouldn't this mean that it should also pass DONT-WAIT nil?

Below that, the docstring mentions what will happen if Emacs has been 
idle for N seconds (N < SECS), but doesn't clarify that if N => SECS, it 
will fire immediately.

>> https://github.com/mooz/js2-mode/commit/b02f4a0d72d0e2087038fe891e2580c4505415ef
>
> Btw, in this line:
>
>    (timer-activate-when-idle timer nil)))
>
> I think you can lose the last argument, since it is optional anyway.

Sure, I wrote it that way just to be explicit.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> Date: Wed, 19 Sep 2012 20:21:32 +0400
> From: Dmitry Gutov <dgutov@yandex.ru>
> CC: jan.h.d@swipnet.se, 12447@debbugs.gnu.org, hanche@math.ntnu.no
> 
> By the way, here's what run-with-idle-timer docstring says:
> "Perform an action the next time Emacs is idle for SECS seconds."
> 
> Shouldn't this mean that it should also pass DONT-WAIT nil?

No, it just means no one considered the possibility that an idle timer
will re-invoke itself like that.  IOW, the doc string is inaccurate.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 19.09.2012 20:38, Eli Zaretskii wrote:
>> Date: Wed, 19 Sep 2012 20:21:32 +0400
>> From: Dmitry Gutov <dgutov@yandex.ru>
>> CC: jan.h.d@swipnet.se, 12447@debbugs.gnu.org, hanche@math.ntnu.no
>>
>> By the way, here's what run-with-idle-timer docstring says:
>> "Perform an action the next time Emacs is idle for SECS seconds."
>>
>> Shouldn't this mean that it should also pass DONT-WAIT nil?
>
> No, it just means no one considered the possibility that an idle timer
> will re-invoke itself like that.  IOW, the doc string is inaccurate.

Okay. Could you fix it?

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Chong Yidong]

Eli Zaretskii <eliz@gnu.org> writes:

>> Date: Wed, 19 Sep 2012 20:21:32 +0400
>> From: Dmitry Gutov <dgutov@yandex.ru>
>> CC: jan.h.d@swipnet.se, 12447@debbugs.gnu.org, hanche@math.ntnu.no
>> 
>> By the way, here's what run-with-idle-timer docstring says:
>> "Perform an action the next time Emacs is idle for SECS seconds."
>> 
>> Shouldn't this mean that it should also pass DONT-WAIT nil?
>
> No, it just means no one considered the possibility that an idle timer
> will re-invoke itself like that.  IOW, the doc string is inaccurate.

I'm not 100% sure this is merely a doc string problem.  In the face of
ambiguity, we should try to choose the behavior that is least likely to
lead to infloops in user code.

When `run-with-idle-timer' is called from an idle timer, we could
interpret it to mean "run the function the next time Emacs becomes idle
for SECS seconds, not including the current period of idleness".

Such behavior seems quite reasonable.  And it is of course easy to
implement, by simply making `run-with-idle-timer' check if it is running
while idle and, if so, giving a nil second arg to
`timer-activate-by-idle'.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> From: Chong Yidong <cyd@gnu.org>
> Cc: Dmitry Gutov <dgutov@yandex.ru>,  hanche@math.ntnu.no,  12447@debbugs.gnu.org
> Date: Thu, 20 Sep 2012 12:04:51 +0800
> 
> Eli Zaretskii <eliz@gnu.org> writes:
> 
> >> Date: Wed, 19 Sep 2012 20:21:32 +0400
> >> From: Dmitry Gutov <dgutov@yandex.ru>
> >> CC: jan.h.d@swipnet.se, 12447@debbugs.gnu.org, hanche@math.ntnu.no
> >> 
> >> By the way, here's what run-with-idle-timer docstring says:
> >> "Perform an action the next time Emacs is idle for SECS seconds."
> >> 
> >> Shouldn't this mean that it should also pass DONT-WAIT nil?
> >
> > No, it just means no one considered the possibility that an idle timer
> > will re-invoke itself like that.  IOW, the doc string is inaccurate.
> 
> I'm not 100% sure this is merely a doc string problem.  In the face of
> ambiguity, we should try to choose the behavior that is least likely to
> lead to infloops in user code.

There's no infloop with my patch.  Assuming that no one comes with a
loop in a couple of days, I will install that, and this particular
issue will be gone.

In general, I think we should prefer solutions that allow Lisp
programmers or users do dumb things (and sometimes get dumb results),
without wedging Emacs.  Restricting what Lisp can do because someone
dumb could wedge Emacs runs a risk of punishing the innocent lot
because of a guilty (or dumb) few.

> When `run-with-idle-timer' is called from an idle timer, we could
> interpret it to mean "run the function the next time Emacs becomes idle
> for SECS seconds, not including the current period of idleness".

I could think of some legitimate uses of the current behavior, though.
For example, imagine an idle timer that gets run after 1 sec of
idleness, does something quick and simple, and then calls itself or
another time with a larger timeout value, at which time it will do
something different, perhaps more complex.  I see no reason to
forcibly prevent such use cases.

It is also in line with Emacs behavior since about forever.  The
possibility of an infloop was a side effect of another bugfix.  With
that possibility hopefully taken care of, we have no reason to change
old behavior, which evidently at least one external package relies
upon.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Chong Yidong]

Eli Zaretskii <eliz@gnu.org> writes:

>> When `run-with-idle-timer' is called from an idle timer, we could
>> interpret it to mean "run the function the next time Emacs becomes idle
>> for SECS seconds, not including the current period of idleness".
>
> I could think of some legitimate uses of the current behavior, though.
> For example, imagine an idle timer that gets run after 1 sec of
> idleness, does something quick and simple, and then calls itself or
> another time with a larger timeout value, at which time it will do
> something different, perhaps more complex.  I see no reason to
> forcibly prevent such use cases.

It would still be possible to get the old behavior by calling
lower-level timer functions used by `run-with-idle-timer'.

> It is also in line with Emacs behavior since about forever.  The
> possibility of an infloop was a side effect of another bugfix.  With
> that possibility hopefully taken care of, we have no reason to change
> old behavior, which evidently at least one external package relies
> upon.

What is the package that relies on the old behavior?

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> From: Chong Yidong <cyd@gnu.org>
> Cc: dgutov@yandex.ru,  hanche@math.ntnu.no,  12447@debbugs.gnu.org
> Date: Fri, 21 Sep 2012 11:31:51 +0800
> 
> What is the package that relies on the old behavior?

js2-mode, see bug #12326 (merged with this one).

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Chong Yidong]

Eli Zaretskii <eliz@gnu.org> writes:

>> What is the package that relies on the old behavior?
>
> js2-mode, see bug #12326 (merged with this one).

I'm confused.  You just said that js2-mode needed a code change:

>> The one I posted in #12326: calling timer-activate-when-idle with nil
>> DONT-WAIT argument.
>
> I don't think that's a workaround.  I think that's what js2 should do,
> if (AFAIU) it wants the timer handler be invoked once, after Emacs has
> been idle for more than 1 sec.

If I understand correctly, js2-mode (prior to the workaround) assumed
the new behavior: it called run-with-idle-timer from inside the idle
timer, with the same delay, with the intention of scheduling for the
next period of idleness.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> From: Chong Yidong <cyd@gnu.org>
> Cc: dgutov@yandex.ru,  hanche@math.ntnu.no,  12447@debbugs.gnu.org
> Date: Fri, 21 Sep 2012 17:09:54 +0800
> 
> Eli Zaretskii <eliz@gnu.org> writes:
> 
> >> What is the package that relies on the old behavior?
> >
> > js2-mode, see bug #12326 (merged with this one).
> 
> I'm confused.  You just said that js2-mode needed a code change:
> 
> >> The one I posted in #12326: calling timer-activate-when-idle with nil
> >> DONT-WAIT argument.
> >
> > I don't think that's a workaround.  I think that's what js2 should do,
> > if (AFAIU) it wants the timer handler be invoked once, after Emacs has
> > been idle for more than 1 sec.
> 
> If I understand correctly, js2-mode (prior to the workaround) assumed
> the new behavior: it called run-with-idle-timer from inside the idle
> timer, with the same delay, with the intention of scheduling for the
> next period of idleness.

I have no idea what js2-mode wants to do.  What I wrote above was
based on my limited reading of the code fragments posted to that bug.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Dmitry Gutov]

On 21.09.2012 13:09, Chong Yidong wrote:
> Eli Zaretskii <eliz@gnu.org> writes:
>
>>> What is the package that relies on the old behavior?
>>
>> js2-mode, see bug #12326 (merged with this one).
>
> I'm confused.  You just said that js2-mode needed a code change:
>
>>> The one I posted in #12326: calling timer-activate-when-idle with nil
>>> DONT-WAIT argument.
>>
>> I don't think that's a workaround.  I think that's what js2 should do,
>> if (AFAIU) it wants the timer handler be invoked once, after Emacs has
>> been idle for more than 1 sec.
>
> If I understand correctly, js2-mode (prior to the workaround) assumed
> the new behavior: it called run-with-idle-timer from inside the idle
> timer, with the same delay, with the intention of scheduling for the
> next period of idleness.

Like I mentioned, the new behavior makes more sense for it, but it 
worked almost the same with the old behavior until the infloop problem.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Chong Yidong]

Eli Zaretskii <eliz@gnu.org> writes:

>> If I understand correctly, js2-mode (prior to the workaround) assumed
>> the new behavior: it called run-with-idle-timer from inside the idle
>> timer, with the same delay, with the intention of scheduling for the
>> next period of idleness.
>
> I have no idea what js2-mode wants to do.  What I wrote above was
> based on my limited reading of the code fragments posted to that bug.

OK, I guess your approach is fine.  Please go ahead and commit your
patch when you're ready, and amend the docstring of run-with-idle-timer
as Dmitry suggested.

Thanks.

bug#12447: 24.1.50; Stuck in garbage collection on OS X

[Eli Zaretskii]

> From: Chong Yidong <cyd@gnu.org>
> Cc: dgutov@yandex.ru,  hanche@math.ntnu.no,  12447@debbugs.gnu.org
> Date: Fri, 21 Sep 2012 22:26:18 +0800
> 
> Eli Zaretskii <eliz@gnu.org> writes:
> 
> >> If I understand correctly, js2-mode (prior to the workaround) assumed
> >> the new behavior: it called run-with-idle-timer from inside the idle
> >> timer, with the same delay, with the intention of scheduling for the
> >> next period of idleness.
> >
> > I have no idea what js2-mode wants to do.  What I wrote above was
> > based on my limited reading of the code fragments posted to that bug.
> 
> OK, I guess your approach is fine.  Please go ahead and commit your
> patch when you're ready, and amend the docstring of run-with-idle-timer
> as Dmitry suggested.

Done in trunk revision 110138.  I'm therefore closing this bug.