Hi,

We found that commits 46bdb4277f98 and 0388f9c74330 (merged in v5.12) introduced more than 40% performance regression in UnixBench's process creation benchmark when comparing 5.10 with 5.15 and 6.1 on AWS' virtualized Graviton instances.

This has been mostly fixed by the following upstream commit:

""" commit a89c6bcdac22bec1bfbe6e64060b4cf5838d4f47 Author: Gabriel Krisman Bertazi krisman@suse.de Date: Mon Jan 9 12:19:55 2023 -0300

arm64: Avoid repeated AA64MMFR1_EL1 register read on pagefault path """ (This is merged in v6.3 and applies cleanly in 5.15 and 6.1).

We performed functional and performance regression tests on both 5.15 and 6.1 with this patch applied. We can also observe 10% improvement in system time in other benchmarks.

Just to be transparent, we initially observed 5% degradation in a benchmark we have with fio randwrite on EBS volumes with our 6.1 downstream kernel but this doesn't seem reproducible and we don't think this patch would cause it.

Here's some quick UnixBench process creation measurements on m7gd.16xlarge instance type. All results are the average of 5 consecutive runs with a single process:

* 5.10.197: 9886.68 lps (baseline) * 6.1.55: 5531.24 lps (unfixed - 44% degradation) * 5.15.133: 5960.12 lps (unfixed - 39% degradation) * 5.15.133: 10306.4 lps (fixed - actually better than baseline) * 6.1.55: 9393.42 lps (fixed - degradation reduced to 5%)

- Luiz