Poor AES performance with rk_crypto2

[tparys]

Just got a ODroid M2 to play with an during initial benchmarking, noticed some weird results. I don't see an M2 specific forum, so @Werner, please move as necessary!

 

Anyway, always interested in crypto throughput, and ARM64 crypto extensions have always put up some good results. So, a little weird when I saw that AES XTS performance seems decent, but CBC seems awful?

 

user@odroidm2:~$ /sbin/cryptsetup benchmark
        aes-cbc        128b       178.5 MiB/s       162.7 MiB/s
        aes-cbc        256b       135.0 MiB/s       211.3 MiB/s
        aes-xts        256b      1399.5 MiB/s      1407.2 MiB/s
        aes-xts        512b      1186.7 MiB/s      1198.7 MiB/s

 

Checking /proc/crypto seems to indicate that rk_crypto2 is being used, so on a whim, I tried doing this:

 

user@odroidm2:~$ sudo rmmod rk_crypto2
[sudo] password for user: 
user@odroidm2:~$ /sbin/cryptsetup benchmark
        aes-cbc        128b       880.4 MiB/s      1722.3 MiB/s
        aes-cbc        256b       755.8 MiB/s      1247.9 MiB/s
        aes-xts        256b      1225.2 MiB/s      1422.5 MiB/s
        aes-xts        512b      1088.9 MiB/s      1123.7 MiB/s

 

Huh.

 

XTS speed is about the same, but CBC is an order of magnitude faster? Might not even be worth using it?

 

Anyway, before I go digging, just curious if anyone's seen this or can reproduce on a Rockchip other than an RK3588S2?

 

Logs from Armbian 26.2.1 (noble/current) -> https://paste.armbian.com/mudojesajo

Edited June 25 by tparys
Added logs

[Werner]

Need to sync tags and boards at some point again...or hopefully this will work automatically at some point.

 

Anyway, try https://github.com/armbian/build/pull/9796 

[tparys]

Good find! Will look through that PR more closely this weekend.

 

Thanks.

 

[Werner]

Let me know. I had hopes the commiter would provide upstreaming effort proof but nothing there yet.