[RFC PATCH v4 03/20] PCI: Make pci_create_root_bus() declare its reliance on MSI domains

Sunil V L sunilvl at ventanamicro.com
Thu Apr 18 04:45:53 PDT 2024


Hi Bjorn,

On Wed, Apr 17, 2024 at 09:03:13PM +0530, Sunil V L wrote:
<snip>
> 
> I think fundamentally there are two issues here.
> 
> 1) MSI domain should have been setup properly when
> pci_register_host_bridge() is called. I see that pci_arch_init() which
> is supposed to get called early calls x86_create_pci_msi_domain().
> pci_register_host_bridge() also calls pci_set_bus_msi_domain() to setup
> the MSI domain which can walk up to host bridge to find. So, not sure
> why PCI_BUS_FLAGS_NO_MSI is getting set. Is there an issue in walking up
> the tree?
> 
I think I understand now why this is happening on X86. In X86, even
though x86_create_pci_msi_domain() sets up the MSI domain, it never
informs IRQ framework. Only later in pcibios_device_add(), the device's
MSI domain will be set which is too late for this bridge logic. So, when
pci_set_bus_msi_domain() is called from pci_register_host_bridge(), it
will not get it.

It works in RISC-V (I guess ARM64 as well) because, the irqchip driver
for MSI controller registers using pci_msi_register_fwnode_provider().
But, if we set bridge->msi_domain = true in X86, it will disable MSI
even though it supports MSI which is really bad.

Please correct me if I am wrong with the analysis. If correct, what do
you think the way forward if using CONFIG option is not good? I tried to
register the domain in below commit but I don't feel it is a cleaner way.
https://github.com/vlsunil/linux/commit/d6cf76f10c225c94fffb286b7a3e803a4e37df54

Thanks,
Sunil



More information about the linux-arm-kernel mailing list