Re: [RFC PATCH] ipc/mqueue: avoid sleep after wakeup

[Manfred Spraul]

On 5/15/21 6:06 AM, Hillf Danton wrote:
> On Fri, 14 May 2021 17:51:47 +0200  Manfred Spraul wrote:
> > On 5/14/21 5:01 AM, Hillf Danton wrote:
> > > The pipeline waker could start doing its job once waiter releases lock and
> > > get the work done before waiter takes a nap, so check wait condition before
> > > sleep to avoid waiting the wakeup that will never come, though that does not
> > > hurt much thanks to timer timeouts like a second.
> > First: The timeout could be infinity, thus the code must not rely on a
> > timeout wakeup.
> >
> > A wrong wait is would be a bug.
> >
> >
> > > Check signal for the same reason.
> > >
> > > Signed-off-by: Hillf Danton <hdanton@xxxxxxxx>
> > > ---
> > >
> > > --- y/ipc/mqueue.c
> > > +++ x/ipc/mqueue.c
> > > @@ -710,15 +710,24 @@ static int wq_sleep(struct mqueue_inode_
> > >    		__set_current_state(TASK_INTERRUPTIBLE);
> > >    
> > >    		spin_unlock(&info->lock);
> > > -		time = schedule_hrtimeout_range_clock(timeout, 0,
> > > -			HRTIMER_MODE_ABS, CLOCK_REALTIME);
> > >    
> > I do not see a bug:
> >
> > We do the __set_current_state() while holding the spinlock. If there is
> > a wakeup, then the wakeup will change current->state to TASK_RUNNING.
> Correct.
> > schedule() will not remove us from the run queue when current->state is
> > TASK_RUNNING. The same applies if there are pending signals: schedule()
> > checks for pending signals and sets current->state to TASK_RUNNING.
> >
> > Since the __set_current_state() is done while we hold info->lock, and
> > since the wakeup cannot happen before we have dropped the lock [because
> > the task that wakes us up needs the same lock], I do not see how a
> > wakeup could be lost.
> >
> > Thus: Which issue do you see?
> 	waiter		waker
> 	----		----
> 	unlock
> 			lock
> 	irq		set STATE_READY
> 	softirq		unlock
> 			wakeup
> 	sleep a tick
> 	  schedule();
>
> No need to schedule given READY.

This is not possible to avoid:

	waiter		waker
	----		----
	unlock
	schedule();
	  calls __schedule()
	   <before rq_lock()>

 	                lock
			set STATE_READY
			unlock
			wakeup
			--> set waiter->state = TASK_RUNNING

Now the run queue will be evaluated even though there is strictly speaking no need to do that.
Changes in ipc/sem.c can't solve that: From what I see, the majority of the critical window is in kernel/sched/*.c and not in ipc/sem.c
I do not consider it as useful to add complexity just to reduce the size of a extremely rare event.

Especially: No harm is done. User space can be preempted at any time, so the kernel is always allowed to check the run queue.

--
	Manfred