Kristian Evensen says:
  In a project I am involved in, we are running ipsec (Strongswan) on
  different mt7621-based routers. Each router is configured as an
  initiator and has around ~30 tunnels to different responders (running
  on misc. devices). Before the flow cache was removed (kernel 4.9), we
  got a combined throughput of around 70Mbit/s for all tunnels on one
  router. However, we recently switched to kernel 4.14 (4.14.48), and
  the total throughput is somewhere around 57Mbit/s (best-case). I.e., a
  drop of around 20%. Reverting the flow cache removal restores, as
  expected, performance levels to that of kernel 4.9.

When pcpu xdst exists, it has to be validated first before it can be
used.

A negative hit thus increases cost vs. no-cache.

As number of tunnels increases, hit rate decreases so this pcpu caching
isn't a viable strategy.

Furthermore, the xdst cache also needs to run with BH off, so when
removing this the bh disable/enable pairs can be removed too.

Kristian tested a 4.14.y backport of this change and reported
increased performance:

  In our tests, the throughput reduction has been reduced from around -20%
  to -5%. We also see that the overall throughput is independent of the
  number of tunnels, while before the throughput was reduced as the number
  of tunnels increased.
Reported-by: Kristian Evensen <kristian.evensen@gmail.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>