On Fri, 31 Oct 2014, Jiang Liu wrote:

On 2014/10/29 5:37, Thomas Gleixner wrote:

quoted

Then it calls down the domain allocation chain. x86_msi_domain would
simply hand down to the parent domain. That would either be the remap
domain or the vector domain.

The issue here is that, the hierarchy irqdomain maintains a tree
topology and every irqdomain only supports one parent.

In case of irq remapping, we need to build one irqdomain for each IOMMU
unit to support hotplug and simplify the implementation. So we need to
build one MSI irqdomain for each IOMMU unit too instead of using a
common MSI irqdomain.

That makes indeed a difference.
 

Current design is that, a common MSI irqdomain to support all MSI when
irq remapping is disabled, and one MSI irqdomain for each IOMMU unit
when irq remapping is enabled.

So we have the code below to choose the correct irqdomain for MSI.
        domain = irq_remapping_get_irq_domain(&info);
        if (domain == NULL)
                domain = msi_default_domain;
        if (domain == NULL)
                return -ENOSYS;

Right. I guess we need to keep it that way for now.

But looking at the code makes me wonder why we actually need to call
into the remap code and do a list walk to figure the domain out. The
association of device and iommu should be known at startup/hotplug
time already. That's out of the scope of this work, but should be
fixed eventually.

Thanks,

	tglx