Re: [PATCH RFC] percpu: add data dependency barrier in percpu accessors and operations

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Jun 20, 2014 at 10:29:04AM -0500, Christoph Lameter wrote:
> On Thu, 19 Jun 2014, Paul E. McKenney wrote:
> 
> > Or just keep doing what I am doing.  What exactly is the problem with it?
> > (Other than probably needing to clean up the cache alignment of some
> > of the per-CPU structures?)
> 
> Writing to a cacheline of another processor can impact performance of that
> other processor since the cacheline (which may contain other performance
> critical data) is evicted from that processors cache.

I believe that most of the people on this thread already understand this,
and that most of them also understand the used of alignment directives
to avoid false-sharing issues.

> The mechanisms for handling percpu data are not designed with the
> consideration of writes into foreign percpu data areas in mind. Surprises
> may result from such use.
> 
> In particular I see a danger in understanding what "atomic" percpu
> operations are. These are not to be confused with regular atomic ops.
> Percpu atomics are atomic for accesses that occur in a single specific
> hardware thread. Percpu "atomics" are atomic vs. interrupts or preemption
> occuring on that specific processor. No serialization is supported for
> accesses may it be read or write from foreign processors.

It sounds like you are thinking strictly in terms of machine-word
sized and aligned per-CPU data.  Much of the cross-CPU accesses are
to structs placed into per-CPU data.  You are not thinking in terms
of having all of the per-CPU data mapped to the same virtual address,
so that CPUs simply cannot access each others' per-CPU data, are you?
That would result in a re-proliferation of NR_CPUS-element arrays.

							Thanx, Paul