Re: [RFC PATCH v4 1/9] CPU hotplug: Provide APIs to prevent CPUoffline from atomic context

[Oleg Nesterov]

(sorry if you see this email twice)

On 12/19, Srivatsa S. Bhat wrote:
>
> On 12/19/2012 01:13 AM, Oleg Nesterov wrote:
>
> >> I tried thinking about other ways to avoid that smp_mb() in the reader,
> >
> > Just in case, I think there is no way to avoid mb() in _get (although
> > perhaps it can be "implicit").
> >
>
> Actually, I was trying to avoid mb() in the _fastpath_, when there is no
> active writer. I missed stating that clearly, sorry.

Yes, I meant the fastpath too,

> > The writer changes cpu_online_mask and drops the lock. We need to ensure
> > that the reader sees the change in cpu_online_mask after _get returns.
> >
>
> The write_unlock() will ensure the completion of the update to cpu_online_mask,
> using the same semi-permeable logic that you pointed above. So readers will
> see the update as soon as the writer releases the lock, right?

Why?

Once again, the writer (say) removes CPU_1 from cpu_online_mask, and
sets writer_signal[0] = 0, this "enables" fast path on CPU_0.

But, without a barrier, there is no guarantee that CPU_0 will see the
change in cpu_online_mask after get_online_cpus_atomic() checks
writer_signal[0] and returns.

Hmm. And this means that the last version lacks the additional rmb()
(at least) if writer_active() == F.

Oleg.

--
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/