Comment # 7 on bug 1127808 from Mel Gorman

(In reply to Jiri Slaby from comment #5)
> We are switching to SLUB, so we can reconsider enabling
> SLAB_FREELIST_HARDENED.

tldr: enable it for master and SLE 15 SP2.

I think it's ok to enable this option based on the preliminary results but it's
borderline at parts. The upstream report that the headline performance impact
is minimal and that is correct but in some cases the results is more variable.
The *average* performance is fine but the outliers are larger. Consider this
slab-intensive workload on UMA for example

http://laplace.suse.de/evaluations/SLE-15-SP2/0002-sec-hardened-freelist/network-netperf-cstate/laurel1/#netperf-udp

Note that the Hmean (harmonic mean) is barely changed but look at the stddev
figures like "Stddev     send-2048" as an example. Relatively speaking it's
much larger. This in itself as the coefficient of variance shows that while it
might be more variable in absoulte terms, it's not much more in relative terms
(varying by 0.93% instead of 0.41%). Because the variance increase is small in
relative terms, I'm not too concerned. It's also possible that this is simply a
co-incidence as FREELIST_HARDENED does not do that much real work as such. 

The vast majority of workloads tested show marginal to no difference. While
some large regressions were reported, they were on NUMA machines for workloads
that occasionally suffer from scheduler decisions (to be researched) and is
*not* related to FREELIST_HARDENED. On that basis, I think it's safe to enable
but we'll keep an eye on it just in case.