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

Hello,
Thu,  8 Sep 2016 16:52:16 +0200, Romain Perier wrote:

quoted

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.

Thank you for your testing. Do you mean that you can reproduce it with
Linux 4.7 too?

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.

It's not easy to detect acceleration by performance because I tested with
Linux 4.7 and older cryptodev (1.8 release) and I got average curl throughput
~22 MiB/s with AES256-SHA256 + cesa. Without cesa I got ~20 MiB/s average with
same configuration.

In any case, I will investigate.

Thank you very much.

Best Regards,
Jussi