Re: [linux-next:master] [crypto] 2f204fe718: stress-ng.af-alg.ops_per_sec 113.7% improvement

Eric Biggers <[email protected]> Wed, 15 Jul 2026 16:27:49 +0000
Newsgroups dev.linux.lists.oe-lkp,org.kernel.vger.linux-crypto,org.kernel.vger.linux-doc
Message-ID <[email protected]>
On Wed, Jul 15, 2026 at 11:00:46PM +0800, kernel test robot wrote:
> 
> 
> Hello,
> 
> 
> we don't have enough knowledge to deeply analyze the impact of this commit
> to stress-ng.af-alg performance tests.
> 
> along with the major kpi improvement:
> 
>    7756377          +113.7%   16578454        stress-ng.af-alg.ops_per_sec
> 
> we also noticed lots of misc tests seem have changes, e.g.
> 
>     697423            -9.5%     630908 ±  2%  stress-ng.af-alg.aes_cipher_ops/sec
>     245103          -100.0%       0.00        stress-ng.af-alg.blowfish_cipher_ops/sec
>     369646          -100.0%       0.00        stress-ng.af-alg.camellia_cipher_ops/sec
>     162192          -100.0%       0.00        stress-ng.af-alg.cast5_cipher_ops/sec
>     115113          -100.0%       0.00        stress-ng.af-alg.cast6_cipher_ops/sec
> 
> we don't know if these are expected behavior upon 2f204fe718.
> so this report is just FYI what we observed in our tests.
> 
> 
> 
> kernel test robot noticed a 113.7% improvement of stress-ng.af-alg.ops_per_sec on:
> 
> 
> commit: 2f204fe718f5bf519013cc2536ad7bb2cbb51661 ("crypto: af_alg - Add af_alg_restrict sysctl, defaulting to 1")
> https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git master

That commit makes AF_ALG block access to some algorithms by default,
overridable by a sysctl.  Apparently these performance tests
misinterpret that as a change in performance.

It actually doesn't make much sense to do performance testing on AF_ALG
at all anymore, given that AF_ALG is deprecated.  A limited form of
AF_ALG is being left in place just for backwards compatibility with a
few programs.  None of those programs use it in a performance-critical
way, as far as I know.

To the extent that it's still done at all, it shouldn't test random
algorithms like "ecb(cast5)" that are obsolete and unused.

- Eric