wait_event_interruptible_timeout() not sleeping accurately

wait_event_interruptible_timeout() not sleeping accurately

[Dick Hollenbeck]

In a user thread I block on wait_event_interruptible_timeout() for about 3300 usecs or
until my threaded interrupt handler calls wake_up_interruptible() when the hardware event
happens.

This seems to work fine for the case when the event actually happens.  But for the timeout
case, the function wait_event_interruptible_timeout() is sleeping 9000 usecs or so rather
than 3300 usecs as requested.  I confirmed this by toggling a GPIO pin before and after
the call and feeding that signal to a logic analyzer.

The user thread is RT round robin at about a priority 5.

Am I doing something disallowed under PREEMPT_RT?
What could be the reason for the inaccurate wait_event_interruptible_timeout() timeout?


			ret = wait_event_interruptible_timeout(bh->rx_waitq,
				bh->rx_intr_flag,
				usecs_to_jiffies(aTimeoutUSecs) );

Re: wait_event_interruptible_timeout() not sleeping accurately

[Sebastian Andrzej Siewior]

On 2019-11-22 13:39:38 [-0600], Dick Hollenbeck wrote:
> 
> In a user thread I block on wait_event_interruptible_timeout() for about 3300 usecs or
> until my threaded interrupt handler calls wake_up_interruptible() when the hardware event
> happens.
> 
> This seems to work fine for the case when the event actually happens.  But for the timeout
> case, the function wait_event_interruptible_timeout() is sleeping 9000 usecs or so rather
> than 3300 usecs as requested.  I confirmed this by toggling a GPIO pin before and after
> the call and feeding that signal to a logic analyzer.
> 
> The user thread is RT round robin at about a priority 5.
> 
> Am I doing something disallowed under PREEMPT_RT?
> What could be the reason for the inaccurate wait_event_interruptible_timeout() timeout?
> 
> 
> 			ret = wait_event_interruptible_timeout(bh->rx_waitq,
> 				bh->rx_intr_flag,
> 				usecs_to_jiffies(aTimeoutUSecs) );

Could you enable some tracing and check what delays the wakeup? There
should be some delay between time timer callback and the wakeup. Maybe
the timer itself is delayed if enough RT tasks are busy.

Sebastian