Re: Enable arm_global_timer for Zynq brakes boot

["Sören Brinkmann" <soren.brinkmann@xilinx.com>]

On Fri, Aug 09, 2013 at 11:32:42AM +0100, Srinivas KANDAGATLA wrote:
> On 08/08/13 18:11, Sören Brinkmann wrote:
> > Hi Daniel,
> > 
> > On Thu, Aug 01, 2013 at 07:48:04PM +0200, Daniel Lezcano wrote:
> >> On 08/01/2013 07:43 PM, Sören Brinkmann wrote:
> >>> On Thu, Aug 01, 2013 at 07:29:12PM +0200, Daniel Lezcano wrote:
> >>>> On 08/01/2013 01:38 AM, Sören Brinkmann wrote:
> >>>>> On Thu, Aug 01, 2013 at 01:01:27AM +0200, Daniel Lezcano wrote:
> >>>>>> On 08/01/2013 12:18 AM, Sören Brinkmann wrote:
> >>>>>>> On Wed, Jul 31, 2013 at 11:08:51PM +0200, Daniel Lezcano wrote:
> >>>>>>>> On 07/31/2013 10:58 PM, Sören Brinkmann wrote:
> >>>>>>>>> On Wed, Jul 31, 2013 at 10:49:06PM +0200, Daniel Lezcano wrote:
> >>>>>>>>>> On 07/31/2013 12:34 AM, Sören Brinkmann wrote:
> >>>>>>>>>>> On Tue, Jul 30, 2013 at 10:47:15AM +0200, Daniel Lezcano wrote:
> >>>>>>>>>>>> On 07/30/2013 02:03 AM, Sören Brinkmann wrote:
> >>>>>>>>>>>>> Hi Daniel,
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> On Mon, Jul 29, 2013 at 02:51:49PM +0200, Daniel Lezcano wrote:
> >>>>>>>>>>>>> (snip)
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> the CPUIDLE_FLAG_TIMER_STOP flag tells the cpuidle framework the local
> >>>>>>>>>>>>>> timer will be stopped when entering to the idle state. In this case, the
> >>>>>>>>>>>>>> cpuidle framework will call clockevents_notify(ENTER) and switches to a
> >>>>>>>>>>>>>> broadcast timer and will call clockevents_notify(EXIT) when exiting the
> >>>>>>>>>>>>>> idle state, switching the local timer back in use.
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> I've been thinking about this, trying to understand how this makes my
> >>>>>>>>>>>>> boot attempts on Zynq hang. IIUC, the wrongly provided TIMER_STOP flag
> >>>>>>>>>>>>> would make the timer core switch to a broadcast device even though it
> >>>>>>>>>>>>> wouldn't be necessary. But shouldn't it still work? It sounds like we do
> >>>>>>>>>>>>> something useless, but nothing wrong in a sense that it should result in
> >>>>>>>>>>>>> breakage. I guess I'm missing something obvious. This timer system will
> >>>>>>>>>>>>> always remain a mystery to me.
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> Actually this more or less leads to the question: What is this
> >>>>>>>>>>>>> 'broadcast timer'. I guess that is some clockevent device which is
> >>>>>>>>>>>>> common to all cores? (that would be the cadence_ttc for Zynq). Is the
> >>>>>>>>>>>>> hang pointing to some issue with that driver?
> >>>>>>>>>>>>
> >>>>>>>>>>>> If you look at the /proc/timer_list, which timer is used for broadcasting ?
> >>>>>>>>>>>
> >>>>>>>>>>> So, the correct run results (full output attached).
> >>>>>>>>>>>
> >>>>>>>>>>> The vanilla kernel uses the twd timers as local timers and the TTC as
> >>>>>>>>>>> broadcast device:
> >>>>>>>>>>> 	Tick Device: mode:     1                                                         
> >>>>>>>>>>> 	Broadcast device  
> >>>>>>>>>>> 	Clock Event Device: ttc_clockevent
> >>>>>>>>>>>
> >>>>>>>>>>> When I remove the offending CPUIDLE flag and add the DT fragment to
> >>>>>>>>>>> enable the global timer, the twd timers are still used as local timers
> >>>>>>>>>>> and the broadcast device is the global timer:
> >>>>>>>>>>> 	Tick Device: mode:     1                                                         
> >>>>>>>>>>> 	Broadcast device                                                                 
> >>>>>>>>>>> 	Clock Event Device: arm_global_timer
> >>>>>>>>>>>
> >>>>>>>>>>> Again, since boot hangs in the actually broken case, I don't see way to
> >>>>>>>>>>> obtain this information for that case.
> >>>>>>>>>>
> >>>>>>>>>> Can't you use the maxcpus=1 option to ensure the system to boot up ?
> >>>>>>>>>
> >>>>>>>>> Right, that works. I forgot about that option after you mentioned, that
> >>>>>>>>> it is most likely not that useful.
> >>>>>>>>>
> >>>>>>>>> Anyway, this are those sysfs files with an unmodified cpuidle driver and
> >>>>>>>>> the gt enabled and having maxcpus=1 set.
> >>>>>>>>>
> >>>>>>>>> /proc/timer_list:
> >>>>>>>>> 	Tick Device: mode:     1
> >>>>>>>>> 	Broadcast device
> >>>>>>>>> 	Clock Event Device: arm_global_timer
> >>>>>>>>> 	 max_delta_ns:   12884902005
> >>>>>>>>> 	 min_delta_ns:   1000
> >>>>>>>>> 	 mult:           715827876
> >>>>>>>>> 	 shift:          31
> >>>>>>>>> 	 mode:           3
> >>>>>>>>
> >>>>>>>> Here the mode is 3 (CLOCK_EVT_MODE_ONESHOT)
> >>>>>>>>
> >>>>>>>> The previous timer_list output you gave me when removing the offending
> >>>>>>>> cpuidle flag, it was 1 (CLOCK_EVT_MODE_SHUTDOWN).
> >>>>>>>>
> >>>>>>>> Is it possible you try to get this output again right after onlining the
> >>>>>>>> cpu1 in order to check if the broadcast device switches to SHUTDOWN ?
> >>>>>>>
> >>>>>>> How do I do that? I tried to online CPU1 after booting with maxcpus=1
> >>>>>>> and that didn't end well:
> >>>>>>> 	# echo 1 > online && cat /proc/timer_list 
> >>>>>>
> >>>>>> Hmm, I was hoping to have a small delay before the kernel hangs but
> >>>>>> apparently this is not the case... :(
> >>>>>>
> >>>>>> I suspect the global timer is shutdown at one moment but I don't
> >>>>>> understand why and when.
> >>>>>>
> >>>>>> Can you add a stack trace in the "clockevents_shutdown" function with
> >>>>>> the clockevent device name ? Perhaps, we may see at boot time an
> >>>>>> interesting trace when it hangs.
> >>>>>
> >>>>> I did this change:
> >>>>> 	diff --git a/kernel/time/clockevents.c b/kernel/time/clockevents.c
> >>>>> 	index 38959c8..3ab11c1 100644
> >>>>> 	--- a/kernel/time/clockevents.c
> >>>>> 	+++ b/kernel/time/clockevents.c
> >>>>> 	@@ -92,6 +92,8 @@ void clockevents_set_mode(struct clock_event_device *dev,
> >>>>> 	  */
> >>>>> 	 void clockevents_shutdown(struct clock_event_device *dev)
> >>>>> 	 {
> >>>>> 	+       pr_info("ce->name:%s\n", dev->name);
> >>>>> 	+       dump_stack();
> >>>>> 	        clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
> >>>>> 	        dev->next_event.tv64 = KTIME_MAX;
> >>>>> 	 }
> >>>>>
> >>>>> It is hit a few times during boot, so I attach a full boot log. I really
> >>>>> don't know what to look for, but I hope you can spot something in it. I
> >>>>> really appreciate you taking the time.
> >>>>
> >>>> Thanks for the traces.
> >>>
> >>> Sure.
> >>>
> >>>>
> >>>> If you try without the ttc_clockevent configured in the kernel (but with
> >>>> twd and gt), does it boot ?
> >>>
> >>> Absence of the TTC doesn't seem to make any difference. It hangs at the
> >>> same location.
> >>
> >> Ok, IMO there is a problem with the broadcast device registration (may
> >> be vs twd).
> > 
> > I have an idea, but no real evidence to prove it:
> > Some of the registers in the arm_global_timer are banked per CPU. I.e.
> > some code must be executed on the CPU the timer is associated with
> > (struct clock_event_device.cpumask) to have the intended effect
> > As far as I can tell, there is no guarantee, that the set_mode()
> > and program_next_event() calls execute on the correct CPU.
> > If this was correct, shutting down the timer for the CPU entering
> > idle might actually shut down the timer for the running CPU, if
> > set_mode() executes on the CPU which is _not_ about to enter idle.
> 
> Hi Sören,
> Am able to reproduce similar issue on StiH415 SOC by enabling both
> global_timer and twd and using cpuidle driver like zynq.
> 
> When CPU0 goes to idle, I noticed that the global timer used for
> boardcast is actually scheduled on wrong cpu.
> My traces for printk like this
> 	printk("DEBUG: %s on CPU:%d CPUMASK:%s\n", __FUNCTION__,
> smp_processor_id(), scpumask);
> 
> shows:
> 
> DEBUG: gt_clockevent_set_mode on CPU:1 CPUMASK: 0
> DEBUG: gt_clockevent_set_next_event on CPU:1 CPUMASK:0
> 
> Which indicates that setting the mode and next_event for a clkevent with
> cpumask 0 is scheduled on cpu1, this will generate an global timer
> interrupt on cpu1 rather than cpu0.
> 
> This might be the reason for cpu0 not coming out of the cpu_idle_loop.

Thanks for reproducing.
I think so too. I had similar debug prints in set_mode and
program_next event, which compared the passed struct clocke_event_device
pointer with the this_cpu_ptr(gt_evt) pointer, which clearly indicate
this mismatch.

	Sören