[crypto] [marvell-cesa] Possible regression after Linux 4.7

[radioconfusion at gmail.com]

Hello,
Fri, 09 Sep 2016 10:24:50 +0300, radioconfusion at gmail.com wrote:
> Thu,  8 Sep 2016 16:52:16 +0200, Romain Perier wrote:
> > > Hello,
> > >
> > > I'm testing the marvell-cesa -driver on Armada 385 board and I think I've found
> > > a regression between Linux 4.7 and 4.8-rc4.
> > > I want to accelerate my curl connections. So I compiled the 4.8-rc4 with
> > > marvell-cesa enabled and HEAD revision of cryptodev-linux.
> > >
> > > Here is my output:
> > >
> > > ~# uname -r
> > > 4.8.0-rc4
> > > ~# modprobe cryptodev
> > > ~# modprobe marvell-cesa
> > > ~# curl -k --ciphers AES256-SHA256 https://myserver/myfile >/dev/null
> > >    % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
> > >                                   Dload  Upload   Total   Spent    Left  Speed
> > >   19  200M   19 39.7M    0     0  16.2M      0  0:00:12  0:00:02  0:00:10 16.2M
> > > curl: (56) SSL read: error:1408F119:SSL routines:SSL3_GET_RECORD:decryption
> > > failed or bad record mac, errno 0
> > >
> --- snip ---
> >
> > Thank you for your feedbacks.
> > I can reproduce the issue with and without multiple engines.

I can still reproduce the issue with Linux 4.9-rc3 (2a26d99b251b8625d27aed14e97fc10707a3a81f)

> > However, I have found an interesting thing, If I use only --ciphers
> > AES256, it works perfectly fine, could you confirm this ?
>
> Yes, I can confirm. And I have found the reason too.
> Most of available ciphers will not be accelerated at all.
> Only these ones I have verified to be accelerated by cesa:
> AES256-SHA256
> AES256-SHA
> AES128-SHA256
> AES128-SHA
>
> When cesa is working, I can see "irq/42-f1090000" on process list of top(1) and
> curl doesn't take all of available cpu.
>
--- snip ---

> > In any case, I will investigate.

Do you have any improvement for the issue?
Please let me know if you need any help to resolve it.

Best Regards,
Jussi