On Wed, Jun 12, 2019 at 02:01:56PM +0100, Marc Zyngier wrote:
> On Wed, 12 Jun 2019 11:17:59 +0100,
> Bharat Kumar Gogada <bharat.kumar.gogada@xxxxxxxxxx> wrote:
> >
> > The current Multi MSI data programming fails if multiple end points
> > requesting MSI and multi MSI are connected with switch, i.e the current
> > multi MSI data being given is not considering the number of vectors
> > being requested in case of multi MSI.
> > Ex: Two EP's connected via switch, EP1 requesting single MSI first,
> > EP2 requesting Multi MSI of count four. The current code gives
> > MSI data 0x0 to EP1 and 0x1 to EP2, but EP2 can modify lower two bits
> > due to which EP2 also sends interrupt with MSI data 0x0 which results
> > in always invoking virq of EP1 due to which EP2 MSI interrupt never
> > gets handled.
>
> I think there is a much simpler explanation for this: Multi-MSI
> mandates that the base interrupt number is naturally aligned to its
> size. Having switches in the middle is just a way to expose the issue,
> but you could see it failing with a single end-point and two MSIs that
> are assigned on an odd boundary.
Agreed, I will rewrite the commit log with a link to the specs,
a switch has no role to play in this bug.
Lorenzo
> > Fix Multi MSI data programming with required alignment by
> > using number of vectors being requested.
> >
> > Fixes: ab597d35ef11 ("PCI: xilinx-nwl: Add support for Xilinx NWL PCIe
> > Host Controller")
> >
> > Signed-off-by: Bharat Kumar Gogada <bharat.kumar.gogada@xxxxxxxxxx>
> > ---
> > V4:
> > - Using a different bitmap registration API whcih serves single and multi
> > MSI requests.
> > ---
> > drivers/pci/controller/pcie-xilinx-nwl.c | 11 +++++------
> > 1 file changed, 5 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/pci/controller/pcie-xilinx-nwl.c b/drivers/pci/controller/pcie-xilinx-nwl.c
> > index 81538d7..a9e07b8 100644
> > --- a/drivers/pci/controller/pcie-xilinx-nwl.c
> > +++ b/drivers/pci/controller/pcie-xilinx-nwl.c
> > @@ -483,15 +483,13 @@ static int nwl_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
> > int i;
> >
> > mutex_lock(&msi->lock);
> > - bit = bitmap_find_next_zero_area(msi->bitmap, INT_PCI_MSI_NR, 0,
> > - nr_irqs, 0);
> > - if (bit >= INT_PCI_MSI_NR) {
> > + bit = bitmap_find_free_region(msi->bitmap, INT_PCI_MSI_NR,
> > + get_count_order(nr_irqs));
> > + if (bit < 0) {
> > mutex_unlock(&msi->lock);
> > return -ENOSPC;
> > }
> >
> > - bitmap_set(msi->bitmap, bit, nr_irqs);
> > -
> > for (i = 0; i < nr_irqs; i++) {
> > irq_domain_set_info(domain, virq + i, bit + i, &nwl_irq_chip,
> > domain->host_data, handle_simple_irq,
> > @@ -509,7 +507,8 @@ static void nwl_irq_domain_free(struct irq_domain *domain, unsigned int virq,
> > struct nwl_msi *msi = &pcie->msi;
> >
> > mutex_lock(&msi->lock);
> > - bitmap_clear(msi->bitmap, data->hwirq, nr_irqs);
> > + bitmap_release_region(msi->bitmap, data->hwirq,
> > + get_count_order(nr_irqs));
> > mutex_unlock(&msi->lock);
> > }
> >
> > --
> > 2.7.4
> >
>
> As for the body of the patch:
>
> Suggested-by: Marc Zyngier <marc.zyngier@xxxxxxx>
> Acked-by: Marc Zyngier <marc.zyngier@xxxxxxx>
>
> Thanks,
>
> M.
>
> --
> Jazz is not dead, it just smells funny.
