Re: [PATCH RFC v2 02/18] irq/dev-msi: Add support for a new DEV_MSI irq domain

[Jason Gunthorpe]

On Wed, Aug 05, 2020 at 10:36:23PM +0000, Dey, Megha wrote:
> Hi Jason,
> 
> > From: Jason Gunthorpe <jgg@xxxxxxxxxxxx>
> > Sent: Wednesday, August 5, 2020 3:16 PM
> > To: Dey, Megha <megha.dey@xxxxxxxxx>
> > Cc: Marc Zyngier <maz@xxxxxxxxxx>; Jiang, Dave <dave.jiang@xxxxxxxxx>;
> > vkoul@xxxxxxxxxx; bhelgaas@xxxxxxxxxx; rafael@xxxxxxxxxx;
> > gregkh@xxxxxxxxxxxxxxxxxxx; tglx@xxxxxxxxxxxxx; hpa@xxxxxxxxx;
> > alex.williamson@xxxxxxxxxx; Pan, Jacob jun <jacob.jun.pan@xxxxxxxxx>; Raj,
> > Ashok <ashok.raj@xxxxxxxxx>; Liu, Yi L <yi.l.liu@xxxxxxxxx>; Lu, Baolu
> > <baolu.lu@xxxxxxxxx>; Tian, Kevin <kevin.tian@xxxxxxxxx>; Kumar, Sanjay K
> > <sanjay.k.kumar@xxxxxxxxx>; Luck, Tony <tony.luck@xxxxxxxxx>; Lin, Jing
> > <jing.lin@xxxxxxxxx>; Williams, Dan J <dan.j.williams@xxxxxxxxx>;
> > kwankhede@xxxxxxxxxx; eric.auger@xxxxxxxxxx; parav@xxxxxxxxxxxx;
> > Hansen, Dave <dave.hansen@xxxxxxxxx>; netanelg@xxxxxxxxxxxx;
> > shahafs@xxxxxxxxxxxx; yan.y.zhao@xxxxxxxxxxxxxxx; pbonzini@xxxxxxxxxx;
> > Ortiz, Samuel <samuel.ortiz@xxxxxxxxx>; Hossain, Mona
> > <mona.hossain@xxxxxxxxx>; dmaengine@xxxxxxxxxxxxxxx; linux-
> > kernel@xxxxxxxxxxxxxxx; x86@xxxxxxxxxx; linux-pci@xxxxxxxxxxxxxxx;
> > kvm@xxxxxxxxxxxxxxx
> > Subject: Re: [PATCH RFC v2 02/18] irq/dev-msi: Add support for a new DEV_MSI
> > irq domain
> > 
> > On Wed, Aug 05, 2020 at 07:18:39PM +0000, Dey, Megha wrote:
> > 
> > > Hence we will only have one create_dev_msi_domain which can be called
> > > by any device driver that wants to use the dev-msi IRQ domain to
> > > alloc/free IRQs. It would be the responsibility of the device driver
> > > to provide the correct device and update the dev->msi_domain.
> > 
> > I'm not sure that sounds like a good idea, why should a device driver touch dev-
> > >msi_domain?
> > 
> > There was a certain appeal to the api I suggested by having everything related to
> > setting up the new IRQs being in the core code.
> 
> The basic API to create the dev_msi domain would be :
> 
> struct irq_domain *create_dev_msi_irq_domain(struct irq_domain *parent)
> 
> This can be called by devices according to their use case.
> 
> For e.g. in dsa case, it is called from the irq remapping driver:
> iommu->ir_dev_msi_domain = create_dev_msi_domain(iommu->ir_domain)
> 
> and from the dsa mdev driver:
> p_dev = get_parent_pci_dev(dev);
> iommu = device_to_iommu(p_dev);
> 
> dev->msi_domain = iommu->ir_dev_msi_domain;
> 
> So we are creating the domain in the IRQ  remapping domain which can be used by other devices which want to have the same IRQ parent domain and use dev-msi APIs. We are only updating that device's msi_domain to the already created dev-msi domain in the driver. 
> 
> Other devices (your rdma driver etc) can create their own dev-msi domain by passing the appropriate parent IRq domain.
> 
> We cannot have this in the core code since the parent domain cannot
> be the same?

Well, I had suggested to pass in the parent struct device, but it
could certainly use an irq_domain instead:

  platform_msi_assign_domain(dev, device_to_iommu(p_dev)->ir_domain);

Or

  platform_msi_assign_domain(dev, pdev->msi_domain)

?

Any maybe the natural expression is to add a version of
platform_msi_create_device_domain() that accepts a parent irq_domain()
and if the device doesn't already have a msi_domain then it creates
one. Might be too tricky to manage lifetime of the new irq_domain
though..

It feels cleaner to me if everything related to this is contained in
the platform_msi and the driver using it. Not sure it makes sense to
involve the iommu?

Jason