Initial benchmarks show they're a net loss:

 $ for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor ; do echo performance > $i; done
 $ echo 4096 32000 64 128 > /proc/sys/kernel/sem
 $ ./sembench -t 2048 -w 1900 -o 0

Pre:

 run time 30 seconds 778936 worker burns per second
 run time 30 seconds 912190 worker burns per second
 run time 30 seconds 817506 worker burns per second
 run time 30 seconds 830870 worker burns per second
 run time 30 seconds 845056 worker burns per second

Post:

 run time 30 seconds 905920 worker burns per second
 run time 30 seconds 849046 worker burns per second
 run time 30 seconds 886286 worker burns per second
 run time 30 seconds 822320 worker burns per second
 run time 30 seconds 900283 worker burns per second

So about 4% faster. (!)

cpu_relax() stalls the pipeline, therefore, when used in a tight loop
it has the following benefits:

 - allows SMT siblings to have a go;
 - reduces pressure on the CPU interconnect.

However, cmpxchg loops are unfair and thus have unbounded completion
time, therefore we should avoid getting in such heavily contended
situations where the above benefits make any difference.

A typical cmpxchg loop should not go round more than a handfull of
times at worst, therefore adding extra delays just slows things down.

Since the llist primitives are new, there aren't any bad users yet,
and we should avoid growing them. Heavily contended sites should
generally be better off using the ticket locks for serialization since
they provide bounded completion times (fifo-fair over the cpus).
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/1315836358.26517.43.camel@twinsSigned-off-by: Ingo Molnar <mingo@elte.hu>