On Mon, May 17 2021 at 18:44, Nitesh Lal wrote:

On Mon, May 17, 2021 at 4:48 PM Thomas Gleixner [off-list ref] wrote:

quoted

The hint was added so that userspace has a better understanding where it
should place the interrupt. So if irqbalanced ignores it anyway, then
what's the point of the hint? IOW, why is it still used drivers?

Took a quick look at the irqbalance repo and saw the following commit:

dcc411e7bf    remove affinity_hint infrastructure

The commit message mentions that "PJ is redesiging how affinity hinting
works in the kernel, the future model will just tell us to ignore an IRQ,
and the kernel will handle placement for us.  As such we can remove the
affinity_hint recognition entirely".

No idea who PJ is. I really love useful commit messages. Maybe Neil can
shed some light on that.

This does indicate that apparently, irqbalance moved away from the usage of
affinity_hint. However, the next question is what was this future
model?

I might have missed something in the last 5 years, but that's the first
time I hear about someone trying to cleanup that thing.

I don't know but I can surely look into it if that helps or maybe someone
here already knows about it?

I CC'ed Neil :)

quoted

Now there is another aspect to that. What happens if irqbalanced does
not run at all and a driver relies on the side effect of the hint
setting the initial affinity. Bah...

Right, but if they only rely on this API so that the IRQs are spread across
all the CPUs then that issue is already resolved and these other drivers
should not regress because of changing this behavior. Isn't it?

Is that true for all architectures?

quoted

While none of the drivers (except the perf muck) actually prevents
userspace from fiddling with the affinity (via IRQF_NOBALANCING) a
deeper inspection shows that they actually might rely on the current
behaviour if irqbalanced is disabled. Of course every driver has its own
convoluted way to do that and all of those functions are well
documented. What a mess.

If the hint still serves a purpose then we can provide a variant which
solely applies the hint and does not fiddle with the actual affinity,
but if the hint is useless anyway then we have a way better option to
clean that up.

+1

= 1

Thanks,

        tglx