Re: [PATCH rc4-mm2 2/2] posix-timers: use try_to_del_timer_sync()

[George Anzinger <george@mvista.com>]

Oleg Nesterov wrote:
> George Anzinger wrote:
> 
>>Oleg Nesterov wrote:
>>
>>>This patch removes timer_active/set_timer_inactive which plays with
>>>timer_list's internals in favour of using try_to_del_timer_sync(),
>>>which was introduced in the previous patch.
>>
>>Is there a particular reason for this, like it does not work, for example, or
>>are you just trying to clean up code?
> 
> 
> It's a cleanup, I think that current code is correct.
> 
> 
>>If this currently works, please leave it alone.
> 
> 
> Ok.
> 
> 
>>We also note that this code is the subject of a patch to the RT patch to cover
>>the same issue when softirqs are run from threads and therefor allow
>>posix_timer_fn to be preempted.  (That fix being mainly to expand usage from
>>just SMP to SMP || SOFTIRQ_PREEMPT.)
> 
> 
> I guess you are talking about this patch:
> http://marc.theaimsgroup.com/?l=linux-kernel&m=111566867218576
> 
> 
>>Also, I think that del_timer_sync and friends need to be turned on if soft_irq
>>is preemptable.
> 
> 
> I agree completely.
> 
> 
>>+ * For RT the timer call backs are preemptable.  This means that folks
>>+ * trying to delete timers may run into timers that are "active" for
>>+ * long times.  To help out with this we provide a wake up function to
>>+ * wake up a caller who wants waking when a timer clears the call back.
>>+ * This is the same sort of thing that the del_timer_sync does, but we
>>+ * need (in the HRT case) to cover two lists and not just the one.
>>+ */
>>+#ifdef CONFIG_PREEMPT_SOFTIRQS
>>+#include <linux/wait.h>
>>+static DECLARE_WAIT_QUEUE_HEAD(timer_wake_queue);
>>+#define wake_timer_waiters() wake_up(&timer_wake_queue)
>>+#define wait_for_timer(timer) wait_event(timer_wake_queue, !timer_active(timer))
> 
> 
> I'm not an expert at all, so I may be wrong, but I don't think
> it's a good idea.
> 
> I think it is bad if __run_timers() could be preempted while
> ->running_timer != NULL. This will interact badly with __mod_timer,
> del_timer_sync. I think that __run_timers() should do:
> 
> 	set_running_timer(base, timer);
> 	preempt_disable();
> 	spin_unlock_irq(&base->lock);
> 
> 	timer->function();
> 
> 	set_running_timer(base, NULL);
> 	preempt_enable();
> 	spin_lock_irq(&base->lock);
> 
> What do you think?

First, I think we need to get Ingo in the discussion. :)

Second, the RT patch has been running this way with little problems, save a 
REALLY intense test we (Monta Vista) have run that, from time to time, shows 
this to be a problem in the posix-timer code that is fixed by including 
SOFTIRQ_PREEMPT as well as SMP in the timer ifdefs.

One thing I do see there (in the RT patch) is a change to del_timer_sync to wait 
for the timer call back to complete rather than to loop...
> 
-- 
George Anzinger   george@mvista.com
High-res-timers:  http://sourceforge.net/projects/high-res-timers/