Re: irq_pipeline: likely/unlikely macro usage

On 16.02.26 13:26, Tobias Schaffner wrote:
> Hi all,
> 
> while staring at the irq_pipelining implementation, I noticed that there
> are some locations that use the unlikely macro for the OOB fast path,
> e.g., [1].
> 
> Could someone explain why we are optimizing for the in-band path here?
> Although the in-band path is indeed more likely, I would have expected
> that we would want to optimize for the OOB path anyway, since this is
> where the saved cycles really matter.
> 

Likely/unlikely optimizations are generally tricky. They are hints which
might even be ignored by concrete archs or in concrete situations.

If they lead to consistently higher latency for oob, I would agree that
we should lean to that path instead. However, if they reduce the average
latency for inband and practically do not increase the worst-case for
oob, it might be a better decision to take that performance optimization
because the worst-case latency is what matters for oob more.

But it remains a fair question whether we can answer that question
whenever we added [un]likely... Did you find inconsistencies in their
usage as well?

Jan

> Best,
> Tobias
> 
> [1] https://gitlab.com/Xenomai/linux-dovetail/-/blob/v5.10.y-dovetail/
> arch/x86/kernel/irq_pipeline.c?ref_type=heads#L260
> 

-- 
Siemens AG, Foundational Technologies
Linux Expert Center