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

["Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>]

On 12/14/2012 11:33 PM, Oleg Nesterov wrote:
> On 12/13, Srivatsa S. Bhat wrote:
>>
>> On 12/13/2012 09:47 PM, Oleg Nesterov wrote:
>>> On 12/13, Srivatsa S. Bhat wrote:
>>>>
>>>> On 12/13/2012 12:42 AM, Srivatsa S. Bhat wrote:
>>>>>
>>>>> Even I don't spot anything wrong with it. But I'll give it some more
>>>>> thought..
>>>>
>>>> Since an interrupt handler can also run get_online_cpus_atomic(), we
>>>> cannot use the __this_cpu_* versions for modifying reader_percpu_refcnt,
>>>> right?
>>>
>>> Hmm. I thought that __this_cpu_* must be safe under preempt_disable().
>>> IOW, I thought that, say, this_cpu_inc() is "equal" to preempt_disable +
>>> __this_cpu_inc() correctness-wise.
>>>
>>> And. I thought that this_cpu_inc() is safe wrt interrupt, like local_t.
>>>
>>> But when I try to read the comments percpu.h, I am starting to think that
>>> even this_cpu_inc() is not safe if irq handler can do the same?
>>>
>>
>> The comment seems to say that its not safe wrt interrupts. But looking at
>> the code in include/linux/percpu.h, IIUC, that is true only about
>> this_cpu_read() because it only disables preemption.
>>
>> However, this_cpu_inc() looks safe wrt interrupts because it wraps the
>> increment within raw_local_irqsave()/restore().
> 
> You mean _this_cpu_generic_to_op() I guess. So yes, I think you are right,
> this_cpu_* should be irq-safe, but __this_cpu_* is not.
> 

Yes.

> Thanks.
> 
> At least on x86 there is no difference between this_ and __this_, both do
> percpu_add_op() without local_irq_disable/enable. But it seems that most
> of architectures use generic code.
> 

So now that we can't avoid disabling and enabling interrupts, I was
wondering if we could exploit this to avoid the smp_mb()..

Maybe this is a stupid question, but I'll shoot it anyway...
Does local_irq_disable()/enable provide any ordering guarantees by any chance?
I think the answer is no, but if it is yes, I guess we can do as shown
below to ensure that STORE(reader_percpu_refcnt) happens before
LOAD(writer_signal).

void get_online_cpus_atomic(void)
{
	unsigned long flags;

	preempt_disable();

	//only for writer
	local_irq_save(flags);
	__this_cpu_add(reader_percpu_refcnt, XXXX);
	local_irq_restore(flags);

	//no need of an explicit smp_mb()

	if (__this_cpu_read(reader_percpu_refcnt) & MASK) {
		this_cpu_inc(reader_percpu_refcnt);
	} else if (writer_active()) {
		...
	}

	this_cpu_dec(reader_percpu_refcnt, XXXX);

}

I tried thinking about other ways to avoid that smp_mb() in the reader,
but was unsuccessful. So if the above assumption is wrong, I guess we'll
just have to go with the version that uses synchronize_sched() at the
writer-side.

Regards,
Srivatsa S. Bhat