Re: [patch 21/32] NTB/msi: Convert to msi_on_each_desc()

[Dave Jiang]

On 12/1/2021 3:16 AM, Thomas Gleixner wrote:
> Jason,
>
> CC+ IOMMU folks
>
> On Tue, Nov 30 2021 at 20:17, Jason Gunthorpe wrote:
> > On Tue, Nov 30, 2021 at 10:23:16PM +0100, Thomas Gleixner wrote:
> > > The real problem is where to store the MSI descriptors because the PCI
> > > device has its own real PCI/MSI-X interrupts which means it still shares
> > > the storage space.
> > Er.. I never realized that just looking at the patches :|
> >
> > That is relevant to all real "IMS" users. IDXD escaped this because
> > it, IMHO, wrongly used the mdev with the IRQ layer. The mdev is purely
> > a messy artifact of VFIO, it should not be required to make the IRQ
> > layers work.
> > I don't think it makes sense that the msi_desc would point to a mdev,
> > the iommu layer consumes the msi_desc_to_dev(), it really should point
> > to the physical device that originates the message with a proper
> > iommu ops/data/etc.
> Looking at the device slices as subdevices with their own struct device
> makes a lot of sense from the conceptual level. That makes is pretty
> much obvious to manage the MSIs of those devices at this level like we
> do for any other device.
>
> Whether mdev is the right encapsulation for these subdevices is an
> orthogonal problem.
>
> I surely agree that msi_desc::dev is an interesting question, but we
> already have this disconnect of msi_desc::dev and DMA today due to DMA
> aliasing. I haven't looked at that in detail yet, but of course the
> alias handling is substantially different accross the various IOMMU
> implementations.
>
> Though I fear there is also a use case for MSI-X and IMS tied to the
> same device. That network card you are talking about might end up using
> MSI-X for a control block and then IMS for the actual network queues
> when it is used as physical function device as a whole, but that's
> conceptually a different case.

Hi Thomas. This is actually the IDXD usage for a mediated device passed 
to a guest kernel when we plumb the pass through of IMS to the guest 
rather than doing previous implementation of having a MSIX vector on 
guest backed by IMS. The control block for the mediated device is 
emulated and therefore an emulated MSIX vector will be surfaced as 
vector 0. However the queues will backed by IMS vectors. So we end up 
needing MSIX and IMS coexist running on the guest kernel for the same 
device.

DJ

> > > I'm currently tending to partition the index space in the xarray:
> > >
> > >   0x00000000 - 0x0000ffff          PCI/MSI-X
> > >   0x00010000 - 0x0001ffff          NTB
> > It is OK, with some xarray work it can be range allocating & reserving
> > so that the msi_domain_alloc_irqs() flows can carve out chunks of the
> > number space..
> >
> > Another view is the msi_domain_alloc_irqs() flows should have their
> > own xarrays..
> Yes, I was thinking about that as well. The trivial way would be:
>
>      struct xarray     store[MSI_MAX_STORES];
>
> and then have a store index for each allocation domain. With the
> proposed encapsulation of the xarray handling that's definitely
> feasible. Whether that buys much is a different question. Let me think
> about it some more.
>
> > > which is feasible now with the range modifications and way simpler to do
> > > with xarray than with the linked list.
> > Indeed!
> I'm glad you like the approach.
>
> Thanks,
>
>          tglx