Re: [PATCH ath-next] wifi: ath10k: spread the PCI copy engine interrupts over the CPUs

From: Loic Poulain

Date: Mon Nov 16 2026 - 10:27:09 EST


On Sat, Nov 14, 2026 at 10:05 PM Julius Bairaktaris
<julius@xxxxxxxxxxxxxx> wrote:
>
> On Mon, Nov 2, 2026 at 1:13 PM Loic Poulain wrote:
> > > - ret = pci_enable_msi(ar_pci->pdev);
> > > - if (ret == 0)
> > > + ret = pci_alloc_irq_vectors(ar_pci->pdev, 1, CE_COUNT_MAX + 1,
> > > + PCI_IRQ_MSI);
> >
> > Why not simply using pci_alloc_irq_vectors_affinity()?
>
> irq_calc_affinity_vectors() sizes the block by the possible CPUs: 5
> vectors on the 4-CPU IPQ8074 with the firmware vector as pre_vector.
> msi_setup_msi_desc() rounds that up and enables 8 in the device, so the
> HTT copy engine's vector 6 has no handler: 226 interrupts in 14 s (one
> run) against 36891-37590 with the hints (three runs). A calc_sets
> callback allocates all 8, but group_cpus_evenly() makes at most one
> group per CPU and the rest get irq_default_affinity. v2 takes the whole
> block or one vector, and keeps the hints.

So, from what you say I understand pci_alloc_irq_vectors_affinity() is
the wrong API here because it sizes the allocation around CPUs,
whereas your goal is to size the allocation around device
CE/capability (8 vectors).

Can you mention cfe9011a05a8 ("ath10k: remove MSI range support").
And why we re-introduce MIS support and how is it different from
previous implementation?

Also, you should update the comment/kdoc for
ath10k_pci_per_engine_handler() which says MSI interrupts are not
handled there.

Regards,
Loic