On Tue, Feb 18, 2025 at 08:41:57AM +0100, Ard Biesheuvel wrote: > > And for IPsec, I'd assume that the cryptd fallback is only needed when > TX and RX are competing for the same CPU. Which can happen if the system is handling encrypted data in both directions. Sometimes you do want to have the hardware steer a given flow to the same CPU in both directions. > So for modern systems, I don't think the SIMD helper does anything > useful, and we should just remove it, especially if we can relax the > softirq/preemption rules for kernel SIMD on x86 like I did for arm64. Sure, if we can ensure that SIMD is always useable even in softirq context then we should definitely remove the simd wrapper. But until that happens, suddenly switching from AESNI to the generic C implementation because a system is loaded is not good. Cheers, -- Email: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> Home Page: http://gondor.apana.org.au/~herbert/ PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
