xfrm: add xdst pcpu cache
retain last used xfrm_dst in a pcpu cache.
On next request, reuse this dst if the policies are the same.
The cache will not help with strict RR workloads as there is no hit.
The cache packet-path part is reasonably small, the notifier part is
needed so we do not add long hangs when a device is dismantled but some
pcpu xdst still holds a reference, there are also calls to the flush
operation when userspace deletes SAs so modules can be removed
(there is no hit.
We need to run the dst_release on the correct cpu to avoid races with
packet path. This is done by adding a work_struct for each cpu and then
doing the actual test/release on each affected cpu via schedule_work_on().
Test results using 4 network namespaces and null encryption:
ns1 ns2 -> ns3 -> ns4
netperf -> xfrm/null enc -> xfrm/null dec -> netserver
what TCP_STREAM UDP_STREAM UDP_RR
Flow cache: 14644.61 294.35 327231.64
No flow cache: 14349.81 242.64 202301.72
Pcpu cache: 14629.70 292.21 205595.22
UDP tests used 64byte packets, tests ran for one minute each,
value is average over ten iterations.
'Flow cache' is 'net-next', 'No flow cache' is net-next plus this
series but without this patch.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/include/net/xfrm.h b/include/net/xfrm.h
index e0feba2..afb4929 100644
--- a/include/net/xfrm.h
+++ b/include/net/xfrm.h
@@ -317,6 +317,7 @@ int xfrm_policy_register_afinfo(const struct xfrm_policy_afinfo *afinfo, int fam
void xfrm_policy_unregister_afinfo(const struct xfrm_policy_afinfo *afinfo);
void km_policy_notify(struct xfrm_policy *xp, int dir,
const struct km_event *c);
+void xfrm_policy_cache_flush(void);
void km_state_notify(struct xfrm_state *x, const struct km_event *c);
struct xfrm_tmpl;