Re: [PATCH 2/2] PCI: iproc: Support multi-MSI only on uniprocessor kernel

[Pali Rohár]

On Monday 07 June 2021 09:48:21 Ray Jui wrote:
> On 6/6/2021 5:30 AM, Sandor Bodo-Merle wrote:
> > The interrupt affinity scheme used by this driver is incompatible with
> > multi-MSI as it implies moving the doorbell address to that of another MSI
> > group.  This isn't possible for multi-MSI, as all the MSIs must have the
> > same doorbell address. As such it is restricted to systems with a single
> > CPU.
> > 
> > Fixes: fc54bae28818 ("PCI: iproc: Allow allocation of multiple MSIs")
> > Reported-by: Marc Zyngier <maz@xxxxxxxxxx>
> > Signed-off-by: Sandor Bodo-Merle <sbodomerle@xxxxxxxxx>
> > ---
> >  drivers/pci/controller/pcie-iproc-msi.c | 8 +++++++-
> >  1 file changed, 7 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/pci/controller/pcie-iproc-msi.c b/drivers/pci/controller/pcie-iproc-msi.c
> > index 557d93dcb3bc..81b4effeb130 100644
> > --- a/drivers/pci/controller/pcie-iproc-msi.c
> > +++ b/drivers/pci/controller/pcie-iproc-msi.c
> > @@ -171,7 +171,7 @@ static struct irq_chip iproc_msi_irq_chip = {
> >  
> >  static struct msi_domain_info iproc_msi_domain_info = {
> >  	.flags = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
> > -		MSI_FLAG_MULTI_PCI_MSI | MSI_FLAG_PCI_MSIX,
> > +		MSI_FLAG_PCI_MSIX,
> >  	.chip = &iproc_msi_irq_chip,
> >  };
> >  
> > @@ -250,6 +250,9 @@ static int iproc_msi_irq_domain_alloc(struct irq_domain *domain,
> >  	struct iproc_msi *msi = domain->host_data;
> >  	int hwirq, i;
> >  
> > +	if (msi->nr_cpus > 1 && nr_irqs > 1)
> > +		return -EINVAL;
> > +
> 
> This should never happen since the framework would have guarded against
> this. But I guess it does not hurt to have the check here.

Yes, this should not happen, but I suggested to add a comment or assert
or some other way to document this kind of constrain. Lot of times code
is copy+pasted to new drivers and because only this one driver has
.alloc function which is using nr_cpus for allocating msi bitmap, it
really makes sense to document this constrain also explicitly.

> >  	mutex_lock(&msi->bitmap_lock);
> >  
> >  	/*
> > @@ -540,6 +543,9 @@ int iproc_msi_init(struct iproc_pcie *pcie, struct device_node *node)
> >  	mutex_init(&msi->bitmap_lock);
> >  	msi->nr_cpus = num_possible_cpus();
> >  
> > +	if (msi->nr_cpus == 1)
> > +		iproc_msi_domain_info.flags |=  MSI_FLAG_MULTI_PCI_MSI;
                                              ^^
Just a small note: there are two spaces instead of just one

Otherwise looks good to me:

Acked-by: Pali Rohár <pali@xxxxxxxxxx>

> > +
> >  	msi->nr_irqs = of_irq_count(node);
> >  	if (!msi->nr_irqs) {
> >  		dev_err(pcie->dev, "found no MSI GIC interrupt\n");
> > 
> 
> Looks fine to me. Thanks.
> 
> Acked-by: Ray Jui <ray.jui@xxxxxxxxxxxx>