Re: [RFC net-next] net ipv6: convert fib6_table rwlock to a percpu lock

On Mon, 2017-07-31 at 19:57 -0700, Shaohua Li wrote:

On Mon, Jul 31, 2017 at 04:10:07PM -0700, Stephen Hemminger wrote:

quoted

On Mon, 31 Jul 2017 10:18:57 -0700
Shaohua Li [off-list ref] wrote:

quoted

From: Shaohua Li <redacted>

In a syn flooding test, the fib6_table rwlock is a significant
bottleneck. While converting the rwlock to rcu sounds straighforward,
but is very challenging if it's possible. A percpu spinlock is quite
trival for this problem since updating the routing table is a rare
event. In my test, the server receives around 1.5 Mpps in syn flooding
test without the patch in a dual sockets and 56-CPU system. With the
patch, the server receives around 3.8Mpps, and perf report doesn't show
the locking issue.

Cc: Wei Wang <redacted>

You just reinvented brlock...

you mean lglock? It has been removed from kernel.
 

quoted

RCU is not that hard, why not do it right?

Maybe. But don't think it's the reason why we shouldn't do the percpu lock now,
this is a simple change, if some smart guys find a way of RCU, we can easily
remove this.

Make sure to test this on a 256 cpu host, dealing with ICMP messages a
lot.

percpu locks do not scale. This hack was okay last decade, sure, but it
is no longer a good hack.

I would rather focus on the RCU work, Wei is actively working on it.