AES-NI XTS Crypto Performance Looking Good For AMD With Linux 5.12 Fix

[phoronix]

Phoronix: AES-NI XTS Crypto Performance Looking Good For AMD With Linux 5.12 Fix

Of the performance-related changes with Linux 5.12 worth noting is faster AES-NI XTS performance for systems relying upon return trampolines "Retpolines" as part of the CPU's Spectre V2 mitigations. On the Intel side this primarily impacts older CPUs where Retpolines is still used while on the AMD side through Zen 3 the Retpolines is still relied upon, which as shown by these benchmarks is now much better off for AMD Ryzen AES XTS performance as measured by Cryptsetup...

http://www.phoronix.com/scan.php?page=news_item&px=AES-NI-CTS-Linux-5.12-AMD

[dwagner]

Certainly a welcome change, as running without Spectre mitigations is just irresponsible.
Even more so now that malicious exploit code has been published: https://therecord.media/first-fully-...overed-online/

[Guest]

Wow, that's a large improvement! Makes a real difference for people using AES-XTS encrypted partitions on NVME SSDs.

[torsionbar28]

Would be interesting to see this compared with FreeBSD's GELI disk encryption which also uses AES in XTS mode, how it performs on the same hardware as compared with Linux.

[archola]

Would love to test this myself, but I'm not sure if I want to use 5.12-rc1 yet. Have the data loss issues which have been reported a couple of days ago been fixed already?
Is applying the patches to 5.11 a good idea or do they depend on other stuff from 5.12?

[schlemil]

Is applying the patches to 5.11 a good idea or do they depend on other stuff from 5.12?

I was tempted to try it, but 5.12 has a huge amount of crypto changes:

https://github.com/torvalds/linux/commit/31caf8b2a847214be856f843e251fc2ed2cd1075