Re: [PATCH v3] ring-buffer: Race when writing and swapping cpu buffer in parallel

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Jul 21, 2014 at 08:43:17AM -0700, Paul E. McKenney wrote:
> On Mon, Jul 21, 2014 at 04:43:24PM +0200, Petr Mládek wrote:
> > 2. Go back, do the swap on any CPU, and do memory barriers via IPI.
> > 
> >    I wonder if the needed memory barrier in rb_reserve_next_event()
> >    could be avoided by calling IPI from ring_buffer_swap_cpu().
> > 
> >    I mean that rb_reserve_next_event() will include the current check
> >    for swapped ring buffer without barriers. But
> >    ring_buffer_swap_cpu() will interrupt the affected CPU and
> >    basically do the barrier there only when needed.
> > 
> >    But I am not sure how this is different from calling
> >    smp_call_function_single() from ring_buffer_swap_cpu().
> >    And I am back on the question why it is dangerous with disabled
> >    interrupts. I can't find any clue in git history. And I miss this
> >    part of the picture :-(
> 
> IIRC, deadlock in the case where two CPUs attempt to invoke
> smp_call_function_single() at each other, but both have
> interrupts disabled.  It might be possible to avoid this by telling
> smp_call_function_single() not to wait for a response, but this often
> just re-introduces the deadlock at a higher level.

FWIW, this is what smp_call_function_single_async() does. But then the call
must synchronized such that no concurrent call happen until the IPI completion.

Otherwise you also have irq_work_queue_on() (not yet upstream but in tip/timers/nohz
and tip/sched/core).