mbox series

[net-next,0/4] net/tls: small TX offload optimizations

Message ID 20190907053000.23869-1-jakub.kicinski@netronome.com
Headers show
Series net/tls: small TX offload optimizations | expand

Message

Jakub Kicinski Sept. 7, 2019, 5:29 a.m. UTC 
Hi!

This set brings small TLS TX device optimizations. The biggest
gain comes from fixing a misuse of non temporal copy instructions.
On a synthetic workload modelled after customer's RFC application
I see 3-5% percent gain.

Jakub Kicinski (4):
  net/tls: unref frags in order
  net/tls: use RCU for the adder to the offload record list
  net/tls: remove the record tail optimization
  net/tls: align non temporal copy to cache lines

 net/tls/tls_device.c | 121 ++++++++++++++++++++++++++++++-------------
 1 file changed, 84 insertions(+), 37 deletions(-)

Comments

David Miller Sept. 7, 2019, 4:11 p.m. UTC | #1 
From: Jakub Kicinski <jakub.kicinski@netronome.com>
Date: Fri,  6 Sep 2019 22:29:56 -0700

> Hi!
> 
> This set brings small TLS TX device optimizations. The biggest
> gain comes from fixing a misuse of non temporal copy instructions.
> On a synthetic workload modelled after customer's RFC application
> I see 3-5% percent gain.

Series applied.

But if history is any indication I'd watch for how much this actually
helps or hurts universally.  We once tried to use non-temporal stores
for sendmsg/recvmsg copies and had to turn that off because it only
helped in certain situations on certain cpus and hurt in others.