Re: Is read barrier missed in kfifo?

[Peter Zijlstra]

On Fri, May 11, 2018 at 09:20:53AM -0700, Paul E. McKenney wrote:
> On Fri, May 11, 2018 at 10:32:42AM +0200, Peter Zijlstra wrote:
> > On Fri, May 11, 2018 at 03:25:18PM +0800, Xiao Guangrong wrote:
> > > 
> > > Hi,
> > > 
> > > Currently, there is no read barrier between reading the index
> > > (kfifo.in) and fetching the real data from the fifo.
> > > 
> > > I am afraid that will cause the vfifo is observed as not empty
> > > however the data is not actually ready for read. Right?
> > 
> > That code is decidedly dodgy indeed. I can only see smp_wmb() but no
> > matching barriers at all -- therefore the code is almost certainly as
> > good as not having any barriers at all.
> > 
> > I would suggest you try and convert the code to smp_store_release() and
> > smp_load_acquire() while you're at it.
> 
> Isn't this one of the places where we rely on control dependencies?

Then it bloody well should have a comment. But at least one side of the
fifo needs a read barrier I think. We can rely on a ctrl-dep on the
write side, where we read the head/tail values, compute space and then
conditionally allow writes to happen.

But on the read side it's all reads and ctrl-dep doesn't help anything.