Re: [patch v2 3/4] percpu_counter: fix code for 32bit systems

[Tejun Heo <tj@kernel.org>]

On Wed, Apr 13, 2011 at 03:57:18PM +0800, shaohua.li@intel.com wrote:
>  static inline s64 percpu_counter_read(struct percpu_counter *fbc)
>  {
> +#if BITS_PER_LONG == 32
> +	s64 count;
> +	unsigned long flags;
> +	spin_lock_irqsave(&fbc->lock, flags);
> +	count = fbc->count;
> +	spin_unlock_irqrestore(&fbc->lock, flags);
> +	return count;
> +#else
>  	return fbc->count;
> +#endif

I don't think this is safe.  The possible deadlock scenario was
percpu_counter_read() being called from irq context and adding irq
locking to percpu_counter_read() doesn't change that in any way.  You
should be changing locking in other places.  Given that the next patch
would make this dancing with locks all pointless, my suggestion is to
drop this patch and proceed with atomic64_t conversion directly and
note that the conversion also removes possible 32bit deviation on
32bit archs.

Thanks.

-- 
tejun