On 20.06.2017 23:30, Tejun Heo wrote:
> Hello,
>
> On Tue, Jun 20, 2017 at 11:28:30PM +0300, Nikolay Borisov wrote:
>>> Heh, looks like I was confused. __percpu_counter_add() is not
>>> irq-safe. It disables preemption and uses __this_cpu_read(), so
>>> there's no protection against irq. If writeback statistics want
>>> irq-safe operations and it does, it would need these separate
>>> operations. Am I missing something?
>>
>> So looking at the history of the commit initially there was
>> preempt_disable + this_cpu_ptr which was later changed in:
>>
>> 819a72af8d66 ("percpucounter: Optimize __percpu_counter_add a bit
>> through the use of this_cpu() options.")
>>
>> I believe that having __this_cpu_read ensures that we get an atomic
>> snapshot of the variable but when we are doing the actual write e.g. the
>> else {} branch we actually use this_cpu_add which ought to be preempt +
>> irq safe, meaning we won't get torn write. In essence we have atomic
>> reads by merit of __this_cpu_read + atomic writes by merit of using
>> raw_spin_lock_irqsave in the if() branch and this_cpu_add in the else {}
>> branch.
>
> Ah, you're right. The initial read is speculative. The slow path is
> protected with irq spinlock. The fast path is this_cpu_add() which is
> irq-safe. We really need to document these functions.
>
> Can I bother you with adding documentation to them while you're at it?
Sure, I will likely resend with a fresh head on my shoulders.
>
> Thanks.
>
