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

[Dey, Megha]

Hi Thomas,

On 8/6/2020 10:10 AM, Thomas Gleixner wrote:
> Megha,
>
> "Dey, Megha" <megha.dey@xxxxxxxxx> writes:
>
> > > -----Original Message-----
> > > From: Jason Gunthorpe <jgg@xxxxxxxxxxxx>
> <SNIP>
> > > Subject: Re: [PATCH RFC v2 02/18] irq/dev-msi: Add support for a new DEV_MSI
> > > irq domain
> can you please fix your mail client not to copy the whole header of the
> mail you are replying to into the mail body?

oops, i hope i have fixed it now..

> > > > > Well, I had suggested to pass in the parent struct device, but it
> > > Oops, I was thinking of platform_msi_domain_alloc_irqs() not
> > > create_device_domain()
> > >
> > > ie call it in the device driver that wishes to consume the extra MSIs.
> > >
> > > Is there a harm if each device driver creates a new irq_domain for its use?
> > Well, the only harm is if we want to reuse the irq domain.
> You cannot reuse the irq domain if you create a domain per driver. The
> way how hierarchical domains work is:
>
> vector --- DMAR-MSI
>         |
>         |-- ....
>         |
>         |-- IR-0 --- IO/APIC-0
>         |        |
>         |        |-- IO/APIC-1
>         |        |
>         |        |-- PCI/MSI-0
>         |        |
>         |        |-- HPET/MSI-0
>         |
>         |-- IR-1 --- PCI/MSI-1
>         |        |
>
> The outermost domain is what the actual device driver uses. I.e. for
> PCI-MSI it's the msi domain which is associated to the bus the device is
> connected to. Each domain has its own interrupt chip instance and its
> own data set.
>
> Domains of the same type share the code, but neither the data nor the
> interrupt chip instance.
>
> Also there is a strict parent child relationship in terms of resources.
> Let's look at PCI.
>
> PCI/MSI-0 depends on IR-0 which depends on the vector domain. That's
> reflecting both the flow of the interrupt and the steps required for
> various tasks, e.g. allocation/deallocation and also interrupt chip
> operations. In order to allocate a PCI/MSI interrupt in domain PCI/MSI-0
> a slot in the remapping unit and a vector needs to be allocated.
>
> If you disable interrupt remapping all the outermost domains in the
> scheme above become childs of the vector domain.
>
> So if we look at DEV/MSI as a infrastructure domain then the scheme
> looks like this:
>
> vector --- DMAR-MSI
>         |
>         |-- ....
>         |
>         |-- IR-0 --- IO/APIC-0
>         |        |
>         |        |-- IO/APIC-1
>         |        |
>         |        |-- PCI/MSI-0
>         |        |
>         |        |-- HPET/MSI-0
>         |        |
>         |        |-- DEV/MSI-0
>         |
>         |-- IR-1 --- PCI/MSI-1
>         |        |
>         |        |-- DEV/MSI-1
>
>
> But if you make it per device then you have multiple DEV/MSI domains per
> IR unit.
>
> What's the right thing to do?
>
> If the DEV/MSI domain has it's own per IR unit resource management, then
> you need one per IR unit.
>
> If the resource management is solely per device then having a domain per
> device is the right choice.

Thanks a lot Thomas for this detailed explanation!!

The dev-msi domain can be used by other devices if they too would want 
to follow the
vector->intel IR->dev-msi IRQ hierarchy.
I do create one dev-msi IRQ domain instance per IR unit. So I guess for 
this case,
it makes most sense to have a dev-msi IRQ domain per IR unit as opposed 
to create one
per individual driver..
> Thanks,
>
>          tglx