Re: [Qemu-devel] [RFC patch 0/6] vfio based pci pass-through for qemu/KVM on s390

[Frank Blaschka <blaschka@linux.vnet.ibm.com>]

On Mon, Sep 22, 2014 at 02:47:31PM -0600, Alex Williamson wrote:
> On Fri, 2014-09-19 at 13:54 +0200, frank.blaschka@de.ibm.com wrote:
> > This set of patches implements a vfio based solution for pci
> > pass-through on the s390 platform. The kernel stuff is pretty
> > much straight forward, but qemu needs more work.
> > 
> > Most interesting patch is:
> >   vfio: make vfio run on s390 platform
> > 
> > I hope Alex & Alex can give me some guidance how to do the changes
> > in an appropriate way. After creating a separate iommmu address space
> > for each attached PCI device I can successfully run the vfio type1
> > iommu. So If we could extend type1 not registering all guest memory
> > (see patch) I think we do not need a special vfio iommu for s390
> > for the moment.
> > 
> > The patches implement the base pass-through support. s390 specific
> > virtualization functions are currently not included. This would
> > be a second step after the base support is done.
> > 
> > kernel patches apply to linux-kvm-next
> > 
> > KVM: s390: Enable PCI instructions
> > iommu: add iommu for s390 platform
> > vfio: make vfio build on s390
> > 
> > qemu patches apply to qemu-master
> > 
> > s390: Add PCI bus support
> > s390: implement pci instruction
> > vfio: make vfio run on s390 platform
> > 
> > Thx for feedback and review comments
> 
> Sending patches as attachments makes it difficult to comment inline.
>
Sorry, don't understand this. I sent every patch as separate email so
you can comment directly on the patch. What do you prefer?
 
> 2/6
>  - careful of the namespace as you're changing functions from static and
> exporting them
>  - doesn't seem like functions need to be exported, just non-static to
> call from s390-iommu.c
> 
Ok, will change this.

> 6/6
>  - We shouldn't need to globally disable mmap, each VFIO region reports
> whether it supports mmap and vfio-pci on s390 should indicate mmap is
> not supported on the platform.
Yes, this is even better to let the kernel announce a BAR can not be
mmap'ed. Checking the kernel code I realized the BARs are valid for
mmap'ing but the s390 platform does simply not allow this. So I feal we
have to introduce a platform switch in kernel. How about this ...

--- a/drivers/vfio/pci/vfio_pci.c
+++ b/drivers/vfio/pci/vfio_pci.c
@@ -377,9 +377,11 @@ static long vfio_pci_ioctl(void *device_

                        info.flags = VFIO_REGION_INFO_FLAG_READ |
                                     VFIO_REGION_INFO_FLAG_WRITE;
+#ifndef CONFIG_S390
                        if (pci_resource_flags(pdev, info.index) &
                            IORESOURCE_MEM && info.size >= PAGE_SIZE)
                                info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
+#endif
                        break;
                case VFIO_PCI_ROM_REGION_INDEX:
                {

>  - INTx should be done the same way, the interrupt index for INTx should
> report 0 count.  The current code likely doesn't handle this, but it
> should be easy to fix.
The current code is fine. Problem is the card reports an interrupt index
(PCI_INTERRUPT_PIN) but again the platform does not support INTx at all.
So we need a platform switch as well. 

>  - s390_msix_notify() vs msix_notify() should be abstracted somewhere

Platform does not have have an apic so there is nothing we could emulate
in qemu to make the existing msix_notify() work.

> else.  How would an emulated PCI device with MSI-X support work?
>  - same for add_msi_route
Same here, we have to setup an adapter route due to the fact MSIX
notifications are delivered as adapter/thin IRQs on the platform.

Any suggestion or idea how a better abstraction could look like?

With all the platform constraints I was not able to find a suitable
emulated device. Remember s390:
- does not support IO BARs
- does not support INTx only MSIX
- in reality currently there is only a PCI network card available
- platform does not support fancy I/O like usb or audio :-)
  So we don't even have kernel (host and guest) support for this
  kind of devices.

>  - We can probably come up with a better way to determine which address
> space to connect to the memory listener.
Any suggestion or idea for that?

> 
> Looks like a reasonable first pass, good re-use of vfio code.  Thanks,
> 
> Alex
> 

Thx,

Frank

>