Re: [PATCH RFC] crypto: caam - never share the gcm and skcipher descriptors serially

From: Sahil Malhotra

Date: Mon Nov 16 2026 - 02:58:38 EST


Hi Vincent,

On Fri, Nov 13, 2026 at 01:23:34AM +0100, Vincent Jardin wrote:
> No, I do not have other CPUs than the LX2160 : your feedbacks with i.MX
> is welcomed.

Sure, we will run the same comparison on i.MX platforms and share the
results on this thread.

> Since SERIAL vs NEVER impacts the performance depending on the CPU and
> on the selected crypto algorithm as you suggest, the code should be more
> dynamic, instead of a single option that slows down all the other cases.
>
> How could we keep "selecting" the best one for each case ?

Agreed, it should be dynamic. But I think we first need the numbers
for each platform before choosing how to select. Depending on what they
show, the choice could be per SoC (e.g. derived from CHA_NUM_MS[DECO_NUM]),
per algorithm, or both. Picking the mechanism now would be guesswork.

To keep the numbers comparable, could you share the exact benchmark
setup you used on LX2160A (tool and command line, e.g. tcrypt mode or
openssl/cryptodev, buffer sizes, number of threads/tfms)? We will use
the same on i.MX.

Also, if possible, could you run the same on LX2160A for authenc
(e.g. authenc(hmac(sha256),cbc(aes))) and xts(aes)? They use the same
kind of shared descriptors, so it would help to see whether the gain
is specific to gcm/skcipher or not.

Thanks,
Sahil