Re: [PATCH v3 00/10] Rework REFCOUNT_FULL using atomic_fetch_* operations

[Hanjun Guo]

On 2019/10/7 23:46, Will Deacon wrote:
> Hi all,
> 
> This is version three of the patches I previously posted here:
> 
>   v1: https://lkml.kernel.org/r/20190802101000.12958-1-will@xxxxxxxxxx
>   v2: https://lkml.kernel.org/r/20190827163204.29903-1-will@xxxxxxxxxx
> 
> Changes since v2 include:
> 
>   - Remove the x86 assembly version and enable this code unconditionally
>   - Move saturation warnings out-of-line to reduce image bloat
> 
> Cheers,
> 
> Will
> 
> Cc: Kees Cook <keescook@xxxxxxxxxxxx>
> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> Cc: Elena Reshetova <elena.reshetova@xxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Cc: Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx>
> Cc: Hanjun Guo <guohanjun@xxxxxxxxxx>
> Cc: Jan Glauber <jglauber@xxxxxxxxxxx>
> 

I tested on top of 5.4-rc2 (with Jan's open-read-close file test case), on a 96 CPU
cores ARM64 server, I can see no much difference under 24 cores (each 24 core is
a NUMA node), but +5.9% performance improve on 48 cores and +8.4% for 96 cores.

For the ARM64 arch,

Tested-by: Hanjun Guo <guohanjun@xxxxxxxxxx>

Thanks
Hanjun