Re: [PATCH, RFC, tip/core/rcu] v3 scalable classic RCU implementation

[Manfred Spraul <manfred@colorfullife.com>]

Paul E. McKenney wrote:
> Assuming that the ordering of processing pending irqs and marking the
> CPU offline in cpu_online_mask can be resolved as noted above, it should
> work fine -- if a CPU's bit is clear, we can safely ignore it.  The race
> can be resolved by checking the CPU's bit in force_quiescent_state().
>
> Or am I missing something?
>   
Yes, that would work:
Rule 1: after CPU_DEAD, a cpu is gone. The cpu is quiet, rcu callbacks 
must be moved to other cpus, ...
Rule 2: if a cpu is not listed in cpu_online_mask, then it can be 
considered as outside a read-side critical section.

The problem with rule 2 is that it means someone 
[force_quiescent_state()] must poll the cpu_online_mask and look for 
changes.
I'd really prefer a notifier. CPU_DYING is nearly the correct thing, it 
only has to be moved down 3 lines ;-)
(I want to kill the bitmaps, not add a hierarchical bitmap polling system!)
> It is entirely possible that rcu_try_flip_waitack() and
> rcu_try_flip_waitmb() need to check the AND of rcu_cpu_online_map and
> cpu_online_map.  If this really is a problem (and it might well be),
> then the easiest fix is to check for cpu_is_offline(cpu) in both
> rcu_try_flip_waitmb_needed() and rcu_try_flip_waitack_needed(), and
> that in both versions of both functions.  Thoughts?
>   
I made a mistake, get_online_cpus() stores current, not a cpu number. 
Thus the described race it not possible. Perhaps there are other users 
that could deadlock.
I don't know enough about the preempt algorithm, thus I can't confirm if 
your proposal would work or not.

--
    Manfred