cpu_relax: drop lowlatency, introduce yield

For spinning loops people do often use barrier() or cpu_relax().
For most architectures cpu_relax and barrier are the same, but on
some architectures cpu_relax can add some latency.
For example on power,sparc64 and arc, cpu_relax can shift the CPU
towards other hardware threads in an SMT environment.
On s390 cpu_relax does even more, it uses an hypercall to the
hypervisor to give up the timeslice.
In contrast to the SMT yielding this can result in larger latencies.
In some places this latency is unwanted, so another variant
"cpu_relax_lowlatency" was introduced. Before this is used in more
and more places, lets revert the logic and provide a cpu_relax_yield
that can be called in places where yielding is more important than
latency. By default this is the same as cpu_relax on all architectures.

So my proposal boils down to:
- lowest latency: use barrier() or mb() if necessary
- low latency: use cpu_relax (e.g. might give up some cpu for the other
  _hardware_ threads)
- really give up CPU: use  cpu_relax_yield

PS: In the long run I would also try to provide for s390 something
like cpu_relax_yield_to with a cpu number (or just add that to
cpu_relax_yield), since a yield_to is always better than a yield as
long as we know the waiter.
processor.h: remove cpu_relax_lowlatency

As there are no users left, we can remove cpu_relax_lowlatency.

Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
32 files changed