RE: [PATCH 1/2] genirq: Fix the possible synchronize_irq() wait-forever

[Liu, Chuansheng]

> -----Original Message-----
> From: Thomas Gleixner [mailto:tglx@xxxxxxxxxxxxx]
> Sent: Friday, February 21, 2014 7:11 PM
> To: Liu, Chuansheng
> Cc: linux-kernel@xxxxxxxxxxxxxxx; Wang, Xiaoming
> Subject: RE: [PATCH 1/2] genirq: Fix the possible synchronize_irq() wait-forever
> 
> On Fri, 21 Feb 2014, Liu, Chuansheng wrote:
> > Hello Thomas,
> >
> > > -----Original Message-----
> > > From: Thomas Gleixner [mailto:tglx@xxxxxxxxxxxxx]
> > > Sent: Friday, February 21, 2014 6:34 PM
> > > To: Liu, Chuansheng
> > > Cc: linux-kernel@xxxxxxxxxxxxxxx; Wang, Xiaoming
> > > Subject: RE: [PATCH 1/2] genirq: Fix the possible synchronize_irq()
> wait-forever
> > >
> > > On Fri, 21 Feb 2014, Liu, Chuansheng wrote:
> > > > But feels there is another case which the synchronize_irq waited there
> > > forever,
> > > > it is no waking up action from irq_thread().
> > > >
> > > > CPU0                                  CPU1
> > > > disable_irq()                         irq_thread()
> > > >   synchronize_irq()
> > > >     wait_event()
> > > >      adding the __wait into the queue  wake_threads_waitq
> > > >        test threads_active==0
> > > > 					 atomic_dec_and_test(threads_active) 1 -- > 0
> > > >
> > > waitqueue_active(&desc->wait_for_threads)
> > > >                                   <== Here without smp_mb(),
> CPU1
> > > maybe detect
> > > >                                       the queue is still empty??
> > > >      schedule()
> > > >
> > > > It will cause although the threads_active is 0, but irq_thread() didn't do
> the
> > > waking up action.
> > > > Is it reasonable? Then maybe we can add one smp_mb() before
> > > waitqueue_active.
> > >
> > > I think you have a point there, but not on x86 wherre the atomic_dec
> > > and the spinlock on the queueing side are full barriers. For non-x86
> > > there is definitely a potential issue.
> > >
> > But even on X86, spin_unlock has no full barrier, the following scenario:
> > CPU0                   CPU1
> > spin_lock
> >                        atomic_dec_and_test
> > insert into queue
> > spin_unlock
> >                        checking waitqueue_active
> 
> But CPU0 sees the 0, right?
Not be clear here:)
The atomic_read has no barrier.

Found commit 6cb2a21049b89 has one similar smp_mb() calling before
waitqueue_active() on one X86 CPU.



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/