Re: [PATCH] rcu/nocb: Extend checks for offloaded rdp by migrate_disable

[Thomas Gleixner <tglx@linutronix.de>]

On Wed, Jul 28 2021 at 01:32, Frederic Weisbecker wrote:
> On Tue, Jul 27, 2021 at 09:33:42PM +0200, Thomas Gleixner wrote:
>> On Tue, Jul 27 2021 at 10:23, Paul E. McKenney wrote:
>> > On Tue, Jul 27, 2021 at 06:38:15PM +0200, Sebastian Andrzej Siewior wrote:
>> >> One thing that has been overseen is that a task within a migrate-disable
>> >> region (as on PREEMPT_RT with disabled BH) is fully preemptible but may
>> >> not be migrated to another CPU which should be enough to guarantee that
>> >> rdp remains stable.
>> >> 
>> >> Check also disabled migration of the task if the RCU data pointer is
>> >> from current CPU. Put the whole check within an SMP ifdef block since
>> >> without SMP there are not CPU migrations to worry about (also
>> >> task_struct::migration_disabled is missing).
>> >> 
>> >> Cc: Frederic Weisbecker <frederic@kernel.org>
>> >> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
>> >> ---
>> >> I don't fully understand why the CPU-hotplug lock matters here but this
>> >> is beside the point ;)
>> >
>> > If I remember correctly, any attempt to change the offloaded state
>> > must hold off CPU-hotplug operations.  So if the current thread is
>> > holding off CPU-hotplug operations, no other thread can be doing
>> > an offload or de-offload operation.
>> 
>> It only prevents unplugging of a CPU, but not plugging a CPU.
>
> Hmm, but both _cpu_down() and _cpu_up() do cpus_write_lock().
> What did I overlook?

I meant, that preemption disable does not prevent plugging a CPU, but
the final unplug step is prevented because the stomp machine thread
cannot run. Similar for migrate_disable(). The final unplug step can
only happen when all non-pinned tasks have left the migrate disabled
section.

Thanks,

        tglx