Re: [PATCH 0/5] [GIT PULL] updates for tip/tracing/ftrace

[Frederic Weisbecker <fweisbec@gmail.com>]

On Sat, Mar 21, 2009 at 09:09:56PM +0100, Frederic Weisbecker wrote:
> On Sat, Mar 21, 2009 at 07:18:58PM +0100, Ingo Molnar wrote:
> > 
> > * Frederic Weisbecker <fweisbec@gmail.com> wrote:
> > 
> > > >  [<ffffffff8020c79d>] return_to_handler+0x0/0x73
> > > >  [<ffffffff8029ea13>] rcu_pending+0x2c/0x5e
> > > >  [<ffffffff8020c79d>] return_to_handler+0x0/0x73
> > > >  [<ffffffff8026abef>] update_process_times+0x3c/0x77
> > > >  [<ffffffff8020c79d>] return_to_handler+0x0/0x73
> > > >  [<ffffffff802875dd>] tick_periodic+0x6e/0x70
> > > 
> > > 
> > > Still hanging in the timer interrupt.
> > > I guess it makes the timer interrupt servicing too slow and then
> > > once it is serviced, another one is raised.
> > > 
> > > But the cause is perhaps more complex
> > > 
> > > I think you have had too much hanging of this type. I'm preparing 
> > > a fix that checks periodically if the function graph tracer is 
> > > spending too much time in an interrupt.
> > > 
> > > I guess I could count the number of function executed between the 
> > > irq entry and its exit.
> > > 
> > > That's the best: if we are hanging in an interrupt, it could be 
> > > whatever interrupt and the jiffies could not be progressing so I 
> > > can't rely on time but only on number of functions executed.
> > > 
> > > May be 10000 calls is a good threshold before killing the function 
> > > graph inside an interrupt?
> > 
> > i think the problem isnt even the IRQ handler - but the fact that 
> > the (timer) irq handler gets re-triggered - so all we do is 
> > processing timer IRQs.
> > 
> > Your patch would detect a timer IRQ hanging - but it would not 
> > detect the 'system makes no progress because there's always anoter 
> > pending timer IRQ to execute' situation.
> 
> 
> Ah, you're right.
> 
>  
> > So i think we need a "function trace watchdog" - which kills the 
> > tracer if we do more than 100,000,000 entries since we started the 
> > self-test, or so.
> > 
> > 	Ingo
> 
> 
> The problem is that it can happen also on other contexts than selftests.
> For example with ftrace=function_graph or by simply enabling the tracer
> later.
> 
> Sometimes it can happen during the selftests, sometimes it's only
> revealed by manually enabling it. I just remind another hang that
> you reported earlier and which I half-solved by fixing a pointless
> softirq call...
> 

Well, ok let's do that, it will be a first and good stage on debugging
the graph hangs.
I will write this selftest watchdog and ftrace_dump() once we reach
100,000,000 entries.
It will be very helpful to know what really happens and what can be
optimized in this area.

Concerning this ftrace_dump(), I will tune it to let us decide if we want
to kill all tracing or not. For example, in case of a graph hang, we don't have
to bother about other tracers.

Thanks.