On Wed, Sep 27, 2017 at 11:01:13AM +0200, Thomas Gleixner wrote:

The default for 32bit architectures is that they CANNOT do atomic
fetch/write of 8 bytes.

Indeed.

In addition, even the native word size accesses weren't guaranteed to be
atomic, as we hadn't used the appropriate accessors to prevent the
compiler from messing with things behind our backs.

So instead of forcing that to be implemented in every affected
architecture this wants to be addressed in generic code and those few
32bit architectures which can do atomic double word fetch/write
implement the magic functions to do so.

I've done this in v2 [1], which Tejun has queued [2] as a fix for v4.14.

Thanks,
Mark.

[1] https://lkml.kernel.org/r/1506426112-19826-1-git-send-email-mark.rutland@arm.com
[2] https://lkml.kernel.org/r/20170926144118.GA70381@devbig577.frc2.facebook.com