On Wed, Aug 01, 2018 at 05:18:25PM +0530, Varun Prakash wrote:
> On Wed, Aug 01, 2018 at 08:33:23AM +0200, Hannes Reinecke wrote:
> > On 07/31/2018 05:07 PM, Varun Prakash wrote:
> > > If number of interrupt vectors are more than num_online_cpus()
> > > then pci_alloc_irq_vectors_affinity() assigns cpumask based
> > > on num_possible_cpus() to the remaining vectors because of
> > > this interrupt does not generate for these vectors.
> > >
> > > This patch fixes this issue by using pci_alloc_irq_vectors()
> > > instead of pci_alloc_irq_vectors_affinity().
> > >
> > > Signed-off-by: Varun Prakash <varun@xxxxxxxxxxx>
> > > ---
> > > drivers/scsi/csiostor/csio_isr.c | 4 +---
> > > 1 file changed, 1 insertion(+), 3 deletions(-)
> > >
>
>
> > >
> > > - cnt = pci_alloc_irq_vectors_affinity(hw->pdev, min, cnt,
> > > - PCI_IRQ_MSIX | PCI_IRQ_AFFINITY, &desc);
> > > + cnt = pci_alloc_irq_vectors(hw->pdev, min, cnt, PCI_IRQ_MSIX);
> > > if (cnt < 0)
> > > return cnt;
> > >
> > >
> > Hmm.
> > That patch (and the reasoning leading up to it) really sound dodgy.
> > I'd rather fix the interrupt affinity code to handle this case correctly.
>
> Yes, it is better to fix the affinity assignment code, I posted
> this patch to start a discussion on this issue.
>
> I can confirm that if number of interrupt vectors are more than
> num_online cpus() then pci_alloc_irq_vectors_affinity() assigns
> cpumask based on cpu_possible_mask to the remaining vectors.
>
> Following is the logic for creating affinity masks
>
> struct cpumask *
> irq_create_affinity_masks(int nvecs, const struct irq_affinity *affd)
> {
> ...
> /* Spread on present CPUs starting from affd->pre_vectors */
> usedvecs = irq_build_affinity_masks(affd, curvec, affvecs,
> node_to_cpumask, cpu_present_mask,
> nmsk, masks);
>
> /*
> * Spread on non present CPUs starting from the next vector to be
> * handled. If the spreading of present CPUs already exhausted the
> * vector space, assign the non present CPUs to the already spread
> * out vectors.
> */
> if (usedvecs >= affvecs)
> curvec = affd->pre_vectors;
> else
> curvec = affd->pre_vectors + usedvecs;
> cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
> usedvecs += irq_build_affinity_masks(affd, curvec, affvecs,
> node_to_cpumask, npresmsk,
> nmsk, masks);
> ...
> }
>
> > Thomas, can you help here?
Any comments on this issue?
