Re: [PATCH v2 1/2] genirq: Retain depth for managed IRQs across CPU hotplug

[Brian Norris]

Hi Alex,

On Wed, Jun 11, 2025 at 08:56:40AM +0200, Aleksandrs Vinarskis wrote:
> Yes. Dell XPS 9345 is arch/arm64/boot/dts/qcom/x1e80100.dtsi based,
> and Asus Zenbook A14 is arch/arm64/boot/dts/qcom/x1p42100.dtsi based,
> which is a derivative but has a slightly different PCIe setup. So far
> both laptops would behave in the same ways.

Thanks. So that's what I suspected, a DWC/pcie-qcom PCIe driver, and
seemingly standard NVMe on top. pcie-qcom doesn't seem to do anything
weird regarding MSIs or affinity, so I wonder why I can't reproduce the
same symptoms on other NVMe setups so far. I can see some of the NVMe
queue MSIs left disabled (queue 0 / CPU 0 doesn't hit the same bugs,
since we don't hotplug CPU 0), but operations seem to function OK even
when missing a few queues. Maybe that's an implementation-specific
behavior that exhibits differently depending on the exact disk in
question.

Anyway, I think we've probably honed in on the bugs, so this might just
be a curiosity.

> > Thanks for the testing. I've found a few problems with my proposed
> > patch, and I've come up with the appended alternative that solves them.
> > Could you give it a try?
> 
> Just tested, and it appears to solve it, though I see some errors on
> wakeup that I don't remember seeing before. I will test-drive this
> setup for a day to provide better feedback and confirm if it is
> related to the fixup or not.

That's promising, I think. Do feel free to forward info if you think
there's still a problem though. I'll await your feedback before spinning
patches.

Brian