running hrtimer_start on an already active hrtimer?

running hrtimer_start on an already active hrtimer?

[Jiri Bohac]

Hi,


I came across a strange bug (in a very old kernel) that triggers
the
	BUG_ON(timer->state != HRTIMER_STATE_CALLBACK);
in __run_hrtimer().

The code runs hrtimer_start() on an already started hrtimer. 
Looking at the description of hrtimer_start() it looks
like something that is allowed:
	/**
	 * hrtimer_start - (re)start an hrtimer on the current CPU
	...
	 * Returns:
	 *  0 on success
	 *  1 when the timer was active

Is this really supposed to work?

I think it's not immune to this race condition:

CPU0						CPU1
__run_hrtimer()
   __remove_hrtimer(...HRTIMER_STATE_CALLBACK)
      //clears HRTIMER_STATE_ENQUEUED
   ...
   raw_spin_unlock(&cpu_base->lock);
   restart = fn(timer);
						hrtimer_start()
						   __hrtimer_start_range_ns()
						      //remove_hrtimer() does nothing because
						      //  HRTIMER_STATE_ENQUEUED is not set
						      enqueue_hrtimer()
   raw_spin_lock(&cpu_base->lock);
   ...
   BUG_ON(timer->state != HRTIMER_STATE_CALLBACK);
   // state has HRTIMER_STATE_ENQUEUED set
   


Should __hrtimer_start_range_ns() do something like
hrtimer_cancel - i.e. explicitly check for ...
HRTIMER_STATE_CALLBACK?


Thanks,

-- 
Jiri Bohac <jbohac@suse.cz>
SUSE Labs, SUSE CZ

Re: running hrtimer_start on an already active hrtimer?

[Thomas Gleixner]

On Tue, 5 May 2015, Jiri Bohac wrote:
> Hi,
> 
> 
> I came across a strange bug (in a very old kernel) that triggers
> the
> 	BUG_ON(timer->state != HRTIMER_STATE_CALLBACK);
> in __run_hrtimer().
> 
> The code runs hrtimer_start() on an already started hrtimer. 
> Looking at the description of hrtimer_start() it looks
> like something that is allowed:
> 	/**
> 	 * hrtimer_start - (re)start an hrtimer on the current CPU
> 	...
> 	 * Returns:
> 	 *  0 on success
> 	 *  1 when the timer was active
> 
> Is this really supposed to work?
> 
> I think it's not immune to this race condition:
> 
> CPU0						CPU1
> __run_hrtimer()
>    __remove_hrtimer(...HRTIMER_STATE_CALLBACK)
>       //clears HRTIMER_STATE_ENQUEUED
>    ...
>    raw_spin_unlock(&cpu_base->lock);
>    restart = fn(timer);
> 						hrtimer_start()
> 						   __hrtimer_start_range_ns()
> 						      //remove_hrtimer() does nothing because
> 						      //  HRTIMER_STATE_ENQUEUED is not set
> 						      enqueue_hrtimer()
>    raw_spin_lock(&cpu_base->lock);
>    ...
>    BUG_ON(timer->state != HRTIMER_STATE_CALLBACK);
>    // state has HRTIMER_STATE_ENQUEUED set
>    

That's in the conditional path:

       if (restart != HRTIMER_NORESTART) {
       	  BUG_ON(timer->state != HRTIMER_STATE_CALLBACK);
	  ....

Which was intentional when we implemented hrtimers in the very
beginning. We wanted to enforce that restart from the callback is not
mixed with a start from some other place.


We removed that restriction recently (queued for 4.2 in
tip/timers/core)

> Should __hrtimer_start_range_ns() do something like
> hrtimer_cancel - i.e. explicitly check for ...
> HRTIMER_STATE_CALLBACK?

No, you cannot do anything about it other than lifting the restriction
or preventing the site which handles the hrtimer to start it.

Thanks,

	tglx