[ltt-dev] add_timer_on: in-kernel users _all_ buggy ?

[Mathieu Desnoyers]

* Neil Horman (nhorman at tuxdriver.com) wrote:
> On Thu, Feb 18, 2010 at 11:41:18AM -0500, Mathieu Desnoyers wrote:
> > * Neil Horman (nhorman at tuxdriver.com) wrote:
> > > On Thu, Feb 18, 2010 at 09:28:00AM -0500, Mathieu Desnoyers wrote:
> > > > * Mathieu Desnoyers (compudj at krystal.dyndns.org) wrote:
> > > > > * Thomas Gleixner (tglx at linutronix.de) wrote:
> > > > > > On Tue, 16 Feb 2010, Mathieu Desnoyers wrote:
> > > > > > > > The function is called from an IPI. That's a LTTNG problem, not a RT one.
> > > > > > > 
> > > > > > > I use del_timer in IPI to delete lttng per-cpu timers on all CPUs. I
> > > > > > > have to do this because timers created with add_timer_on are documented
> > > > > > > to be incompatible with del_timer_sync():
> > > > > > > 
> > > > > > >  * Synchronization rules: Callers must prevent restarting of the timer,
> > > > > > >  * otherwise this function is meaningless. It must not be called from
> > > > > > >  * interrupt contexts. The caller must not hold locks which would prevent
> > > > > > >  * completion of the timer's handler. The timer's handler must not call
> > > > > > >  * add_timer_on(). Upon exit the timer is not queued and the handler is
> > > > > > >  * not running on any CPU.
> > > > > > 
> > > > > > Errm. The documentation says: 
> > > > > > 
> > > > > >       "The timer's handler must not call add_timer_on()." 
> > > > > > 
> > > > > > It's not talking about a timer which was initialized with
> > > > > > add_timer_on().
> > > > > > 
> > > > > >  And your per cpu timer handlers have no requirement to call
> > > > > > add_timer_on() simply because add/mod_timer() is requeueing the timer
> > > > > > on the same cpu on which the handler runs. 
> > > > > > 
> > > > > > So the IPI is just a solution for a non existing problem.
> > > > > 
> > > > > Oh, right. Thanks for the explanation. I'll look into moving LTTng to a
> > > > > saner del_timer_sync() scheme to delete the timers.
> > > > 
> > > > Double-checking this:
> > > > 
> > > > add_timer_on() needs to be paired with mod_timer_pinned(), otherwise
> > > > NO_HZ SMP config can rebalance the timer to a different CPU. I am fixing
> > > > this in lttng 0.194. These per-cpu timers, of course, should usually be
> > > > deferrable (they are in lttng).
> > > > 
> > > > (looking at kernel 2.6.32.4 here)
> > > > Looking at the kernel/time/clocksource.c watchdog, I wonder how
> > > > del_timer manages to synchronize the timer teardown. The handler,
> > > > clocksource_watchdog(), uses add_timer_on(), which prohibits using
> > > > del_timer_sync(). This seems rather odd. If we remove the watchdog and
> > > > re-add it, it may still be in use while we initialize the timer
> > > > structure.
> > > > 
> > > > Also, net/core/drop_monitor.c trace_drop_common usage of add_timer_on
> > > > seems odd:
> > > > 
> > > > Executing (AFAIK) with preempt on, data points to a per-cpu timer:
> > > > 
> > > >        if (!timer_pending(&data->send_timer)) {
> > > >                 data->send_timer.expires = jiffies + dm_delay * HZ;
> > > >                 add_timer_on(&data->send_timer, smp_processor_id());
> > > >        }
> > > > 
> > > > How is timer_pending synchronized with the target CPU timer wheel ?
> > > > 
> > > Hm, I think I see your point here.  You're suggesting that a call to one of the
> > > tracepoint hooks in the drop monitor can race against a second call to the hook
> > > from an interrupt context that pre-empted the first, leading to double add of
> > > the timer?
> > 
> > Yes.
> > 
> > > I agree, in fact I think its likely worse that that, the shared data
> > > on the skb that I modify there can get corrupted in that case as well.  I expect
> > > a bit of refactoring paired with a local_irq_save/restore should fix that.
> > 
> > For this current code, you don't seem to need to delete the timer. I think
> > you'll have to find a way to do the timer_pending check and the add_timer_on
> > operations atomically wrt timer wheel execution on the CPU (which can be a
> > remote cpu because preemption is enabled). Your case is a bit weird..  the
> > typical scenario here would be to use add_timer_on directly to re-arm the timer
> > for a future expiration time (removing the current pending timer
> > unconditionnally at the same time). But you leave the timer at its current
> > expiration time if present, and only if not present add it. I wonder if it's
> > really your intent ?
> > 
> 
> Yes, thats my intent.  The idea is that when I note a packet is lost, I record
> that in a data buffer that is part of an skb, and schedule a timer to send that
> skb a second or so later.  Other drops in that time frame are recorded within
> the same skb.  Since the drop monitor isn't built as a module currently, I don't
> need to worry about deleting the timer at all, I can just let it expire.
> 
> Sounds like I need to wrap up the data extract and time modifications in a
> prempt_disable/enable and local_irq_save/restore set
> Neil

I fear this will probably break the RT kernel, because the timer spinlocks
become sleepable (I had the same issue within LTTng). Another approach you could
try is to add a new member to the timer API, which hold the local timer base
lock while:

if the timer is already pending
  - waits for already executing timers to complete their execution.
  - add_timer_on local cpu with new expiration.
else
  - do nothing

Thanks,

Mathieu

> 
> > Thanks,
> > 
> > Mathieu
> > 
> >