Re: [PATCH v11 19/19] futex: Allow to make the private hash immutable.

[Shrikanth Hegde <sshegde@linux.ibm.com>]

On 4/10/25 20:58, Sebastian Andrzej Siewior wrote:
> On 2025-04-10 20:22:08 [+0530], Shrikanth Hegde wrote:
>> Hi Sebastian.
> Hi Shrikanth,
> 
>> It would be good option for the application to decide if it needs this.
> 
> You mean to have it as I introduced it here or something else?
> 

as you have introduced it here.

>> Using this option makes the perf regression goes away using previous number of buckets.
> 
> Okay, good to know. You test this on on ppc64le?

Yes.

> 
>> Acked-by: Shrikanth Hegde <sshegde@linux.ibm.com>
>>
>> base:
>> ./perf bench futex hash
>> Averaged 1556023 operations/sec (+- 0.08%), total secs = 10  <<-- 1.5M
>>
>> with series:
>> ./perf bench futex hash -b32768
>> Averaged 126499 operations/sec (+- 0.41%), total secs = 10   <<-- .12M
>>
>> ./perf bench futex hash -Ib32768
>> Averaged 1549339 operations/sec (+- 0.08%), total secs = 10  <<-- 1.5M
> Thank you for testing.
> …
>> nit: Does it makes sense to split this patch into futex and perf?
> 
> First I wanted to figure if we really do this. I have no idea if this
> regression would show up in real world use case or just here as part of
> the micro benchmark.
> And if we do this, it would probably make sense to have one perf patch
> which introduces -b & -I.  And then figure out if the additional option
> to prctl should be part of the resize patch or not. Probably we should
> enforce 0/1 of arg4 from the beginning so maybe folding this in makes
> sense.
>

ok.

> Sebastian