Re: [v2] timers: Fix usleep_range() in the context of wake_up_process()

From: Guenter Roeck
Date: Wed Oct 12 2016 - 12:54:15 EST


On Wed, Oct 12, 2016 at 09:27:35AM -0700, Doug Anderson wrote:
> Hi,
>
> On Wed, Oct 12, 2016 at 9:03 AM, Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> > drivers/iio/accel/kxcjk-1013.c: kxcjk1013_runtime_resume()
> > drivers/iio/accel/bmc150-accel-core.c:bmc150_accel_runtime_resume()
> > drivers/iio/accel/mma8452.c:mma8452_runtime_resume()
> > drivers/iio/accel/mma9551_core.c:mma9551_sleep()
>
> As far as I can tell these drivers will not suffer unduly from my
> change. Worse case they will delay 20us more, which is listed as the
> max.
>
20 ms.

> Also note that I assume the reason you flagged these is because they
> follow the pattern:
>
> if (sleep_val < 20000)
> usleep_range(sleep_val, 20000);
> else
> msleep_interruptible(sleep_val/1000);
>
Correct

> I will note that usleep_range() is and has always been
> uninterruptible, since the implementation says:
>
> void __sched usleep_range(unsigned long min, unsigned long max)
> {
> __set_current_state(TASK_UNINTERRUPTIBLE);
> do_usleep_range(min, max);
> }
>
Good point.

> So I'm not at all convinced that we are changing behavior here. The
> "interruptible" vs. "uninterruptible" affects whether signals can
> interrupt the sleep, not whether a random wake up of a task can. What
> we really need to know is if they are affected by a wakeup.
>
Yes, you are correct.

> > kernel/trace/ring_buffer.c:rb_test()
>
> I assume that the person who wrote this code was confused since they wrote:
>
> set_current_state(TASK_INTERRUPTIBLE);
> /* Now sleep between a min of 100-300us and a max of 1ms */
> usleep_range(((data->cnt % 3) + 1) * 100, 1000);
>
> That doesn't seem to make sense given the first line of usleep_range().
>
... which, for those who don't pay attention (like me), is
__set_current_state(TASK_UNINTERRUPTIBLE);

> In any case, again I don't think I am changing behavior.
>
> > A possible solution might be to introduce usleep_range_interruptible()
> > and use it there.
>
> This could be a useful function, but I don't think we need it if we
> find someone who needs a wakeup to cut short a sleep. We can just
> call one of the schedule functions directly and use a timeout.
>

Agreed.

>
> Thank you for searching through for stuff and for your review, though!
>
No problem. Thanks for correcting me.

Note that I also searched for use of usleep_range() in conjunction with a
a task wakeup, but did not find anything. I did find a large number of cases,
though, where the explicit assumption is made that the minimum sleep time
is well defined.

Guenter